Machine learning-based uncertainty analysis in power system planning: Insights and pathways for decarbonization

Abstract

This study performs an extensive uncertainty analysis of decarbonization pathways for the Canadian power system. By leveraging machine learning (ML) techniques, we represent complexities and nonlinear behaviors of the power system, which are challenging to observe using traditional modeling methods due to computational limitations. We further utilize K-means clustering and statistical methods to identify correlations, investigate the effects of uncertainty in key parameters (capital costs, demand, and carbon tax), and identify opportunities for diversification of the power generation mix. Key findings include the increased flexibility of gas combined cycle (generation mean 4.8 TWh, installed capacity mean 4.8 GW) over gas simple cycle (generation mean 22.3 GWh, 0 installed capacity) in response to carbon tax, highlighting its role in a low-carbon future. Variations in solar and wind deployment rates, influenced by capacity factors and location-specific costs, underscore the necessity of a flexible grid system. This analysis reveals important correlations between renewable deployment, and transmission expansion and gas combined cycle reliance. The results also highlight minimal sensitivity of hydroelectric power generation (∼ 43 % of scenarios) to changes in input parameters. The novelty of our work lies in proposing a comprehensive approach which uses ML techniques to go beyond the capabilities of existing models, specifically by performing extensive uncertainty analysis and provides key insights into system behavior, causative factors, and an expanded breadth of potential pathways. Taken together, this proposed approach lays a foundation for strategic planning and policy formulation.