Urban heat island and electrical load estimation using machine learning in metropolitan area of rio de janeiro

Abstract

We developed a daily electrical load forecasting model for the State of Rio de Janeiro and a monthly model for each Light concessionaire substation in the Metropolitan Area of Rio de Janeiro (MARJ). The data used are 1) daily National System Operator (ONS) electrical load data respecting to State of Rio de Janeiro for four years (2017-2020); 2) the monthly electrical load of 84 Light substations for 11 years (2010-2020); 3) the maximum, minimum, and mean air temperature. In addition, remotely sensed landsurface temperature (LST) based on Landsat data from 1984 to 2020 is used to restructure the current meteorological network on MARJ based on the disposition of the Light substations. Using crossvalidation, we performed 500 daily and 500 monthly training-testing experiments of ve regressive machine learning-based algorithms. Results for daily-ONS and monthly-Light loads show average correlations (hindcast in parentheses) of the tted models of 0.85 ± 0.09 (0.83 ± 0.07) and 0.89 ± 0.05 (0 .91 ± 0.06), respectively. The model's Mean Average of Error (MAE) values correspond to a percentage error of about 4.03% (daily) and 4.83% (monthly). According to the monthly electrical load behavior revealed, when the temperature changes from 23 to 26℃ at MARJ, it rises roughly from 1.92x10 ± 67227.4 kWh to 2.70x10 ± 90198.5 kWh. We performed a cluster analysis based on the locations of 1) the 18 meteorological stations currently installed, 2) the 84 Light electrical load substations, and 3) the urban heat island cores. Results reveal seven locations where new meteorological stations should be installed to model the electrical load with higher spatial resolution in MARJ.