A model for predicting household cooking energy sources using multi-layer perceptron technique

Abstract

Cooking with dirty energy is one of the causes of CO₂ emission, and adversely affects climate change. In Nigeria, most households are using dirty sources of energy for cooking such as firewood, charcoal and kerosene due to availability and cheapness. The aim of the research is to develop a Machine Learning (ML) model to determine households’ cooking energy sources. In this study, a Multi-layer Perceptron (MLP) model was developed to classify households' cooking energy sources in Nigeria. Secondary dataset obtained from the National Bureau of Statistics (NBS) database for Nigerian General Household Panel Survey (NGHPS) was used. Dataset were pre-processed, and Forward Sequential Feature Selection technique was used to select the most relevant features. A dataset consisting a total of 4,980 households was used to develop MLP Architecture by varying the number of hidden layers (1 and 2 layers), and the number of units were selected in multiples of 10 and 5 for 1 and 2 layers respectively. The best accuracy of 96.89% was obtained using 1 –hidden layer at (60) units in the hidden layer, and 97.09% at (70, 70) units using 2-hidden layer architecture. In general, the study shows the significance of using 2-hidden layers architecture and the influence from the use of variations in the number of units in the hidden layers to find the optimal number of units for the problem under question. The study shows that given some specific household's characteristics, the household's Major Cooking Fuel Source in Nigeria can be determine with an accuracy of 97%, precision of 95%, recall and F1 Score of 94% each. In future, this study can be improved by exploring other Machine Learning Techniques for better performance.