Mohi-Ud-din, G (2020) Deep Learning based Approaches for Cost Effective Short-term Energy Load Forecasting And Consumer Behaviour Modelling in Households. Doctoral thesis, Liverpool John Moores University.

Preview

Text 2020GhulamMohiUdDinPhD.pdf - Published Version Download (1MB) | Preview

Abstract

Today, there is a lot of enthusiasm to fulfil global energy needs from alternative energy resources. Due to the increasing demand for electricity, the traditional electricity market relies on decisions to plan electricity systems, and to generate and distribute electricity to their consumers to balance demand and supply. The peak demands of electricity highly affect these decisions and often cause system failure and shortage of electricity. By predicting energy requirements, these peak demands and the uncertainties in human behaviour in households are optimised to balance the load through various demand response programmes. A smart grid ecosystem requires intelligent Home Energy Management Systems (HEMSs) to profile highly non-stationary and non-linear measurements and conduct correlations of such measurements with diverse inputs (e.g. environmental factors) in order to improve the end-user experience, as well as to aid the overall demand-response optimisation process. The huge amount of energy consumption information collected from the individual appliances opens up lots of opportunities to mine the hidden patterns in the data in order to understand the human behaviour related to energy usage. However, processing huge amounts of information for analysis purposes demands lots of resources e.g. time and computational power. Parallelisation techniques allow the processing of large amounts of data while requiring less computational time. However, neural networks widely used in data processing are highly complex to parallelise during the model parallelisation due to their sequential nature. A key challenge here is to exploit the parallelisation capabilities of hardware as well as software in terms of multicore/multithreaded CPUs and GPUs (Graphical Processing Units). To overcome these challenges, in this research work, we propose a new unified approach to predict day, week and month-wide energy consumption by reducing the computation resources and model human behaviour in households in order to save scarce energy resources and improve demand response programmes. We go beyond current profiling schemes by proposing Deep COLA; a Deep Competitive Learning Algorithm that addresses limitations of high dimensional data and enables accurate modelling of appliance-level energy consumption. We show that our proposed scheme is far more computationally efficient and scalable data-wise than three conventional clustering approaches namely, K-Means, DBSCAN and SOM, using real household datasets. This research work includes a number of contributions. The first is a dominant feature selection algorithm from the pool of features to increase the performance of the forecasting model. The second is a prediction model based on deep learning algorithms to improve the forecasting accuracy and processing time. The third is a clustering algorithm based on competitive learning to profile day, week and month-wide energy consumption patterns, using appliance-level data for a given household. The current methods are based on K-Means, DBSCAN and rule mining which require expert knowledge to get improved clustering. However, the proposed concept of competitive learning allows to extract compact and well separated clusters. This approach automatically extracts the optimal number of clusters without using Elbow, Silhouette or Bootstrap methods commonly seen in the existing work. The fourth is the profiling of appliance-level energy consumption in synergy with environmental factors in order to reveal per-household behavioural characteristics under three associations: appliance-to-appliance, appliance-to-time and appliance-to-environment. The last one is a parallelisation approach in the form of data parallelisation to forecast energy consumption by utilizing large amounts of data.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Deep neural networks; Home Energy Management Systems; Competitive learning; Load profiling; Deep learning; Load forecasting
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions:	Computer Science & Mathematics
Date Deposited:	17 Sep 2020 19:20
Last Modified:	16 Sep 2022 00:50
DOI or ID number:	10.24377/LJMU.t.00013652
Supervisors:	Shi, Q, Marnerides, A and Dobbins, C
URI:	https://researchonline.ljmu.ac.uk/id/eprint/13652

View Item