Modeling and Analysis of Electricity Load Forecasting Using Machine learning Approaches

Abstract

Electricity load forecasting has become an important issue due to variations in the operational efficiency of power systems. The load forecasting problem is a regression

problem and can be modeled as a time series problem. Due to the non-linear na- ture of the time series problem and the variable weather conditions, an accurate load

forecast is a difficult undertaking. The accuracy of load forecasting is important for utility companies as well as consumers. In this study, we compare several machine learning algorithms for processing time series data on power use in order to acquire relevant insights. The primary focus of our study is on doing an analysis of the various time series modeling methodologies in order to anticipate future power consumption. There are various types of load forecasting however, our research focus is short-term load forecasting. Forecasting the short-term load is an essential component of many different types of energy optimization, including peak shaving and cost escape. The accumulated load forecasting of a region or a city has been the subject of research for many years and has generated accurate findings. But the building level hasn’t gotten much attention because its dynamics are very different compared to those of a utility or some other medium- or large-scale customers. This thesis study concentrates on making short-term load forecasts at the building level. The correlation technique is applied in order to filter out features that are redundant. Autocorrelation and partial autocorrelation methods are used to locate previous hours that are comparable to the current one. Linear regression (LR), k-Nearest Neighbours (KNN), Moving Average

(MA), Vector Auto-regressive (VAR), Seasonal Auto-Regressive Integrated Moving Av- erage with eXogenous factor (SARIMAX) and Long short-term memory (LSTM) are

applied for forecasting. The data used for the modeling was collected every half an

xii

hour on the electrical load by the Smart Agri Tech lab, SINES, NUST. The estimated forecasting performance demonstrates that LSTM works better, but it takes longer to train. The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the LSTM model are 8179.69, 5380.37, and 7.43%

respectively. This work can be used to forecast the electric load of multi-storey build- ings and will help the managers of the buildings to manage electric resources and reduce

electricity costs.