Please use this identifier to cite or link to this item:
Title: Development of Smart On-device based Electricity Consumption Forecast Model for Energy Management
Alexander Nnamdi Ndife
Yodthong Mensin
ยอดธง เม่นสิน
Naresuan University. School of Renewable Energy and Smart Grid Technology
Keywords: Forecasting
Deep Learning
Ensemble method
Neural Networks
Edge Computing
Real-Time Control
Power Consumption
Issue Date: 2022
Publisher: Naresuan University
Abstract: This thesis developed a Deep Learning algorithm for power consumption forecasting implementable in low memory storage and energy systems like our smartphones, IPad and Tablets. Power forecasting is a multidisciplinary task that forms an important aspect of energy generation, distribution, and management. When forecasting is done in a smart way, it will help in both energy conservation and resource planning. Interestingly, the traditional means of estimating power usage through previous utility bills is nowadays being replaced with machine intelligence. This research is motivated by the quest to determine power consumption expectation for a medium-term (a week ahead) given the current load demand and the possibility of monitoring and controling energy usage at home or offcie in a real-time from anywhere in the world. This forecast model leveraged on multivariate dataset to make a multi-step time series (7 consecutive days ahead) forecast. It is split into: 1) Problem Framing - here, considerations were made on what type of forecast we really want: short, medium or long term; 2) Modelling - it entails finding the consumption behaviour that captures features of the medium-term forecast that was chosen; 3) Forecasting - predicting the power need for the chosen medium-term period of 7days; 4) Smart Home Energy Management System - finding a way to control appliances that consumes energy in order to limit the energy usage within the estimated threshold of the forecast result.This forecast model is based on ConvLSTM-Encoder-Decoder algorithm explicitly designed to enhance the quality of spatiotemporal encodings throughout its feature learning process. However, randomness and other challenges of training neural networks necessitated the ensemble approach used, where multiple models were trained but allowed contribution to prediction by each model to be weighted proportionally to their level of trust and estimated performance on the model. This architecture in principle investigated power consumption of manually operated home against a smart home and its performance tested on time-domain Household Electricity Power Consumption dataset from France; and further validated using a real time load profile collated from School of Renewable Energy and Smart Grid Technology (SGtech), Naresuan University Smart Office. RMSE of 361kWh was recorded compared with 465kWh on persistence model and an improved RMSE of 358kWh was achieved when validated using holdout validation data from the automated office. However, overall performance on error, forecast time and computational speed was later compared with research efforts in literature and the result obtained showed a significant improvement. And comparative analysis carried out between the energy consumption of a manually operated office and a smart office using the proposed SmartHEMS showed that this smart app saved about 24% of the energy normally consumed in the office.
Description: Doctor of Philosophy (Ph.D.)
ปรัชญาดุษฎีบัณฑิต (ปร.ด.)
Appears in Collections:วิทยาลัยพลังงานทดแทนและสมาร์ตกริดเทคโนโลยี

Files in This Item:
File Description SizeFormat 
61031799.pdf4.04 MBAdobe PDFView/Open

Items in NU Digital Repository are protected by copyright, with all rights reserved, unless otherwise indicated.