Abstract

Environmental protection is being progressively considered as paramount condition for the planet's continued habitability. After the Kyoto Protocol signature in 1997, governments, industry stakeholders and academia began to work on the development of effective and efficient environmentally driven policies and economic mechanisms, and the proper design of such parameters is critically dependent on carbon emissions projections. In such scenario, inaccurate carbon emissions predictions may be one of the root factors leading to the overall ineffectiveness of the European Union environmental regulatory framework.
Therefore, the present thesis introduces a novel computational learning framework for carbon emissions prediction incorporating a RReliefF driven features selection and an iterative neural network architecture improvement. Our learning framework algorithmic architecture iteratively chains the features selection process and the backpropagation artificial neural network (NN/BP) architecture design based on the data assessment accomplished by the RReliefF algorithm. Thus a better features set - NN/BP architecture combination is obtained for each specific prediction target.
The implemented framework was trained and validated with real world data obtained from the European Union (Eurostats), the International Energy Agency, the Organization for Economic Co-operation and Development, and the World Bank. The validation dataset comprised 26 potential predictors covering the period 1990 - 2014. Additionally, a case study was conducted with a new dataset comprising data obtained from the World Resources Institute's Climate Data Explorer (CAIT), and the World Bank database. The case study dataset comprises 24 potential predictors covering the period 1970 - 2014.
The learning framework also features an Explainable Artificial Intelligence (XAI) module that provides explanations of the predictions in terms of global features impact and local features weights. The global model explanations are computed by means of partial dependence functions, while local model explanations are computed by means of the interpretable model-agnostic explanations (LIME) algorithm.
The framework evaluation against current mainstream machine learning models, and its benchmarking comparing to recent published researches on carbon emissions prediction indicates that our research contribution is relevant and capable of supporting the improvement of environmental policies. The learning framework outcomes are also expected to provide some ground for future researches targeting carbon emissions causality analysis, as well as potential improvements on both ANNs and XAI techniques.