Files
Abstract
A microgrid is a localized group of electricity sources and loads that normally operates connected to and synchronous with traditional wide area synchronous grid(macrogrid), but can also disconnect to island mode and then function autonomouslyas physical or economic conditions dictate. In this way, a microgrid can effectivelyintegrate various sources of distributed generation (DG), especially Renewable Energy Sources (RES) and can supply emergency power, changing between island andconnected modes. With the various devices connected on the microgrid, there is a lotof useful data which can be used to analyze their individual behaviours as well as forunderstanding how they operate with each other. Data Analytics has made tremendous progress in this decade and is only going to grow. It is natural that the utilitiesindustry will want to make use of this technology of Machine Learning available attheir disposal. This thesis investigates the novel approach of Deep Learning over traditional or statistical data analysis by comparing their performance in the analyzingthe behaviour of the microgrid. Deep Learning architectures like Recurrent NeuralNetworks (RNN), Long Short Term Memory Networks (LSTM) have been proved tobe effective in understanding time series data or temporal sequences in previous research in comparison to statistical algorithms like Auto Regressive Integrated MovingAverage (ARIMA) or Vector Auto Regression (VAR). A new deep learning architecture of a ConvolutionalLSTM (Convolutional Long Short Term Memory Networks)is developed and tested. Finally the results are compared based on time requiredto train the models to their best performance and their accuracy in detecting theanomalies in the microgrid.