COMPARATIVE ANALYSIS BETWEEN ARIMA AND LSTM MODELS IN SHORT-TERM FORECAST OF ACTIVE POWER DEMAND

Abstract

In Brazil, a problem related to electricity consumption for a user connected to the distribution system, in medium and high voltage, is demand overrun. This problem occurs when the measured demand exceeds the contracted demand by more than 5%. How to evaluate the efficiency of a computational model as a solution to the short-term forecast problem of active power demand, aiming its use as a control method for application in demand controllers? This study aims to evaluate the efficiency of the ARIMA statistical model and the LSTM deep neural network model as control methods. Based on historical measurement data, steps were developed to adjust, forecast and evaluate the models in a case study of a medium voltage user connected to the distribution system of the Enel/RJ concessionaire. It was possible to observe that the ARIMA model obtained an efficiency of 46%, that is, 46% of the predicted values ​​with a maximum percentage variation of 5% (degree of precision) of the measured values, while the LSTM model obtained 13%.