Electricity Load Forecasting using Long Short-Term Memory

Abstract

Nowadays, the demand for electric energy is increasing rapidly while the mismatch between electricity demand, production, and distribution is growing. This leads to continuous power outages, which in turn affect the end customers. On the other hand, a sudden change in electricity flow affects electricity power distribution stations and other electricity distribution infrastructures. Thus, knowing electricity demand at a specific time and the nature of how electricity demand changes beforehand help to properly plan and manage future demands. Electricity load forecasting also help electric energy purchasing, transmission, and distribution planning, for demand-side management operation and maintenance. Many previous studies focused on external factors for electricity load forecasting. Collecting this kind of data requires various types of sensors, which increases the cost of time and resources. But also, in some cases, these types of data may not exist at all. Electricity load forecasting can be done at various levels ranging from generation to end users. Previous research in substation load forecasting used an Artificial Neural Network, which does not capture temporal dependencies between data points unless it is specified explicitly. The dataset we used contains 79,969 records of historical electric power load data measured every 15 minutes in the period from 1/1/2020 to 4/13/2022 from Gofa substation in Addis Ababa. This paper proposes Long Short-Term Memory architecture to forecast substation load demand for Gofa substation. The main objective of this study is to improve performance of electricity load forecasting at substation level using Long Short-Term Memory. For the purpose of comparison, various deep learning models are implemented. Mean Squared Error and Root Mean Squared Error are used to test the accuracy of the models. The result shows that the Long Short-Term Memory model outperforms other deep learning models such as Artificial Neural Network and Recurrent Neural Network, which have been used in previous works