Optimisation or Learning? Appropriate Choice of Models and Methods for Sequential Decision‐Making in Power System Operation

Abstract

Power system operation is inherently a sequential decision‐making problem, traditionally addressed through optimisation models built upon high‐quality load forecasts. However, the rapid proliferation of renewable generation and price‐responsive demands, such as electric vehicles and merchant energy storage, has introduced multiple sources of uncertainty that are difficult to characterise and predict with sufficient accuracy. To address the impact of uncertainty, advanced optimisation and reinforcement learning have emerged as two representative yet conceptually distinct paradigms for sequential decision‐making in power systems. Despite their different theoretical foundations, modelling assumptions and implementation schemes, their practical application domains often overlap, which has led to persistent ambiguity in model selection. This paper argues that the choice between optimisation and learning should not be framed as a methodological preference, but as a modelling decision rooted in available information and the structure of uncertainty. This article clarifies the fundamental strengths and limitations of the two paradigms and identifies the conditions under which each is most appropriate. The goal is not to advocate one paradigm over the other, but to establish a principled view of model choice for sequential decision‐making in power system and energy internet operations.