HYBRID MACHINE LEARNING MODELS FOR ENHANCING AIR QUALITY AND ENERGY EFFICIENCY IN SMART CITIES

Abstract

Rapid urbanization poses significant challenges to sustainable development and the living standards of city residents. A smart city makes use of Information and Communication Technologies (ICT) to increase the quality of services given to its citizens in terms of health, transportation, electricity and environment. Due to technological advancement in Internet of Things and machine learning models, smart city applications and research has surged significantly. Smart cities have evolved into true "data engines," continually producing and consuming data. A wide range of IoT devices and applications serve as data sources, recording different everyday activities from all over the place and producing a massive volume of data. The data generated in smart cities is one of the important assets for the deployment of smart cities. The researchers have applied machine learning models on the data collected to establish new smart city era. A significant amount of study using machine learning has been done on smart cities such as smart healthcare, intelligent transportation, smart energy, etc. The efficiency of the machine learning models is an important aspect. It is worth developing an efficient machine learning models for electric energy and air quality application to provide better quality life in smart cities. The major objectives of the present work are as follows: (i) Systematic study and illustration of data driven electricity theft detection methods & its limitations, consequences of electricity theft, datasets, evaluation parameters, and research challenges. (ii) (iii) (iv) Development of hybrid data balancing and machine learning model to detect electricity theft in smart grids. Development and exploration of machine learning models for short-term electricity consumption forecasting at daily and hourly level in residential buildings. Analysis of various machine learning, feature selection, and data imputation techniques to understand the working, performance and finding the best model for PM2.5 prediction. This thesis is organized into six chapters. Chapter 1: Introduction The introduction contains general concept of smart city and its components. It includes general data analytics framework for smart cities, applications of data analytics in smart cities, research challenges, research objectives, and contributions of the thesis. Chapter 2: In this chapter, different data driven methods (machine learning, deep learning, and hybrid methods) and its limitations in detection of electricity theft are discussed. The datasets related to electricity theft detection are presented with metadata like number of instances, sampling frequency, duration, and country. Various performance parameters used by researchers to evaluate the electricity theft detection models with formulas, definition, and the studies which used these parameters is presented. The research challenges and consequences related to electricity theft is also discussed. Chapter 3: This chapter presents the developed hybrid model of electricity theft detection in smart grid which combines the concept of hybrid data balancing and hybrid machine learning models. It presents the comparative analysis proposed approach with state-of-the-art methods. A statistical method (ANOVA) used to validate the results statistically is also presented in this chapter. Chapter 4: This chapter presents the proposal and implementation of machine learning models and their hybrid for short term electricity consumption forecasting at residential level. This chapter contains two parts; in the first part, voting ensemble model optimized with genetic algorithm is designed to forecast electricity consumption and compared with state-of-the-art results and in the second part stacking ensemble model is designed and compared with state of-the-art results. Chapter 5: This chapter presents exploratory analysis of feature selection, data imputation, and models for prediction of PM2.5 pollutant in the air pollution. This chapter contains two parts, in first part, ten regression models, three data imputation techniques are explored on the air pollution dataset and ensemble of best models are performed. In second part, five feature selection models, six regression algorithms are explored to find the best combination to predict PM2.5. Chapter 6: This chapter concludes the thesis and presents the future scope of the work and research directions.