8th World Climate Change Congress

Biography

Biography: Bismark Ameyaw

Abstract

Carbon dioxide emissions (CO₂), a major greenhouse gas, has contributed massively towards the weight of each country’s share in global warming and climate change including the United States (US). The Energy Information Administration (EIA) uses the National Energy Modelling System (NEMS) to forecast and make projections on electric sector CO₂ emissions in the US for short, medium and long-term timelines but the forecast inaccuracies of past projections are considerably high. The massive tradeoffs of factors, unrealized assumptions and scenarios on deterministic and peripheral variables, a uniform time effect used, and volatilities in patterns of electricity consumption are among the factors that cause such high forecast errors. Here we propose and apply a long short-term memory (LSTM) recurrent neural network (RNN) technique devoid of exogenous variables and their assumption and/or scenarios thereof; that allows for varying time effects and capable of withstanding volatilities; to forecast electric power sector CO₂ emissions in the US for the 2005-2015 period. Based on the high-accuracy forecasts, we propose optimal emissions-mitigation pathways to achieve US’s set targets in the electric sector. The empirical results suggest the proposed LSTM model presents overwhelming improvement on accuracy of forecasts and projections of selected EIA's Annual Energy Outlooks (AEOs). We find that the forecasts accuracy on emissions from policy-targeted variables (PTVs), particularly petroleum products, is affected by the use of much outdated historic data as such cause outliers and that using recent reliable data in making short and medium-term forecasts of PTVs is recommended.