Using Deep Learning to Fill Data Gaps in Environmental Footprint Accounting

Environmental footprint accounting relies on economic
input–output
(IO) models. However, the compilation of IO models is costly and time-consuming,
leading to the lack of timely detailed IO data. The RAS method is
traditionally used to predict future IO tables but suffers from doubts
for unreliable estimations. Here we develop a machine learning-augmented
method to improve the accuracy of the prediction of IO tables using
the US summary-level tables as a demonstration. The model is constructed
by combining the RAS method with a deep neural network (DNN) model
in which the RAS method provides a baseline prediction and the DNN
model makes further improvements on the areas where RAS tended to
have poor performance. Our results show that the DNN model can significantly
improve the performance on those areas in IO tables for short-term
prediction (one year) where RAS alone has poor performance, R2 improved from 0.6412 to 0.8726, and median
APE decreased from 37.49% to 11.35%. For long-term prediction (5 years),
the improvements are even more significant where the R2 is improved from 0.5271 to 0.7893 and median average
percentage error is decreased from 51.12% to 18.26%. Our case study
on evaluating the US carbon footprint accounts based on the estimated
IO table also demonstrates the applicability of the model. Our method
can help generate timely IO tables to provide fundamental data for
a variety of environmental footprint analyses.