Parameterizing open-source energy models: Statistical learning to estimate unknown power plant attributes

Energy systems models are used to perform energy and environmental policy analysis, inform company strategy, and understand implications of technological change. Although open-source models can promote transparency and reproducibility, data availability and cost can be prohibitive barriers for researchers and other stakeholders. This paper presents a novel application of a statistical approach to predict unknown power plant parameters in Canada using available data from the United States, which can be applied in other settings where critical model inputs are missing. We apply two statistical learning methods, linear regression and k-nearest-neighbors, and compare their performance on unseen portions of the United States data before applying the learned functions to unknown Canadian data. Results indicate that reasonable predictions of heatrates and, to a lesser extent, operation and maintenance costs are possible even with limited data about age, capacity, and power plant types. The nearest-neighbor approach generally outperforms linear regressions for the datasets and applications to power plant parameters investigated here.