Short-Term Energy Demand Forecasting in Fernando de Noronha with LSTM Neural Networks

Abstract

Given the growth in demand for electricity and the need to expand the matrix through the insertion of new sources such as renewable energy and storage, the planning and operation of the isolated system of Fernando de Noronha Island, it becomes essential to understand the behavior of electricity demand and the variables that influence it. In addition to the need to make forecasts, to ensure safe and adequate operation of the system. Thus, the objectives of this study were to know and extract useful information from the time series of energy demand, as well as the construction of machine learning models for forecasting short-term demand. The work was developed in four stages: 1) Exploratory analysis of the three-year data series of energy demand on an hourly scale, where its statistics were studied and feature engineering was applied to evaluate exogenous variables. 2) Construction of the long short term memory (LSTM) and multi layer perceptron (MLP) predictive models that will serve as a benchmark for comparison. 3) Training and validation of the models to define the hyperparameters; 4) Calculation of the metrics and comparison of the models and errors. It was possible to identify interesting characteristics of energy demand, such as its trend and seasonality, the low climate correlation, and the main factors that influence its behavior. In addition, the LSTM model outperformed the MLP by up to 37% with the MAPE metric of up to 2.26% between forecast periods.