A Reinforcement Learning Approach to Energy Transition Planning Under Uncertainty

Long-term energy transitions require large investments in emerging technologies whose costs are uncertain and evolve with deployment. Investment decisions that fail to account for this uncertainty risk locking in expensive or inefficient pathways if actual technology costs diverge from projections. Existing approaches often rely on expert elicitation of future cost or scenario analysis to address uncertainty, overlooking the adaptive capacity of planners to revise decisions based on the observed evolution of technology costs. In addition, conventional multi-stage optimization methods are computationally constrained, limiting the frequency with which decisions can be revisited and the scale of the system modeled. Our study addresses this gap by proposing a reinforcement learning based framework that develops adaptive policies serving as decision rules that determine technology deployment over time in response to current system conditions, including technology costs that evolve endogenously with deployment. We first embed a stochastic version of Wright’s Law into an energy system model to capture uncertainty in how increased deployment reduces costs over time. We then train deployment policies to minimize expected total system cost across stochastic cost trajectories using direct policy search, a reinforcement learning approach, to model the co-evolutionary feedback between investment and cost. We demonstrate our method using a global energy system model that includes decisions for 15 competing technologies, spanning fossil fuels and renewables, over the period 2020 to 2050. Preliminary results show that adaptive decision-making can reduce system costs across a wide range of future scenarios compared to exogenous, fixed pathways. Our analysis reveals diverse energy expansion pathways that emerge in response to technological evolution, demonstrating how adaptive policies can support more resilient and cost-effective transitions under uncertainty.

Speaker: Mofan Zhang, Stanford University

Register at weblink

Room 101