Integrating Machine Learning with Multi-Horizon Stochastic Programming for Scalable Capacity Expansion

Abstract

We introduce a novel approach to tackling large-scale capacity expansion problems in power systems by integrating machine learning techniques with multi-horizon stochastic programming that combines long-term strategic decision-making under uncertainty with short-term operational decisions and uncertainties. Our proposed method leverages a novel neural network model as a surrogate for the expected operational costs of a given long-term capacity investment decision. To enhance generalization and accuracy, we also develop a new dimension reduction algorithm that reduces the input vector dimensionality of the surrogate model by extracting only a subset of important features of the input decision vectors, derived by analyzing solutions to the deterministic problem, and by downscaling the operational uncertainties typically represented as time series. The approach is demonstrated using EMPIRE, a capacity expansion planning model for Europe. Experimental results indicate that our method achieves a substantial reduction in computation time while also achieveing lower expected costs compared to existing stochastic optimization methods. These findings highlight the effectiveness of our approach in addressing key challenges in stochastic capacity planning. Furthermore, this advancement has the potential to inform future studies in areas such as infrastructure development and resource allocation across various industries.