Probabilistic commodity price projections for unbiased techno-economic analyses

Abstract

Techno-economic analysis is a core methodology for assessing the feasibility of new technologies and processes. The outcome of an analysis is largely dictated by the product's price, as selected by the practitioner. Representative future price distributions are required as inputs to investment, sensitivity, and uncertainty analyses across the 20 to 25 year plant life. However, current price selection procedures are open to subjective judgment, not adequately considered, or neglected by calculating a minimum selling price. This work presents a machine learning methodology to produce unbiased projections of future price distributions for use in a techno-economic analysis. The method uses an ensemble of 100 neural network models with Long Short-Term Memory layers. The models are trained on the Energy Information Administration's (EIA) long-term crude oil projections and a commodity's historic price data. The proposed method is demonstrated by projecting the price of five commodity chemicals 26 years into the future using 12 years of historic data. Alongside the economic outlook extracted from the EIA projections, the five commodity price distributions capture stochastic and deterministic elements specific to each commodity. A statistically significant difference was observed when using the price projections to revise the Net Present Value distributions for two previous techno-economic analyses. This suggests that relying on heuristics when selecting price ranges and distributions is unrepresentative of a commodity's price uncertainty. The novelty of this work is the presentation of an unbiased machine learning approach to project long-term probabilistic prices for techno-economic analyses, emphasising the pitfalls of less rigorous approaches.