Design and development of Residential Sector Load Prediction model during COVID-19 Pandemic using LSTM based RNN

Abstract

Covid-19 pandemic and resulting lockdown has created a wide impact on social life, including sudden rise in residential load demand. Utilities, for better load scheduling and economic operations, rely on different prediction models among which neural networks proved to be more appropriate. For such unforeseen situations, the non-availability of prior predictions elevated the utility challenges. Moreover, the stringency of lockdowns caused due to mutated COVID-19 virus, necessitates accurate lockdown load predictions. This paper proposes a Recurrent Neural Network based Long Short-Term Memory (RNN-LSTM) model, trained to produce such predictions for two areas of residential sector. The model uses real-time residential load data from the year 2020, with and without weather parameters. The correlation factor (R) of proposed method 0.9683 outperformed the ARIMA's value 0.703. The model is evaluated with correlation factors of 0.9683 and 0.9235 without temp; 0.90361 and 0.913662 with temperature for Apurupa and Jyothi colonies respectively located in Hyderabad, India. In addition, the error metrics namely, Mean absolute percentage error (MAPE) and Mean absolute error (MAE) are 2.0464 and 138.576 for Apurupa colony; 0.015 and 201.648 for Jyothi colony respectively. However, the prediction error metrics increased slightly with temperature data. The proposed framework will assist utilities for effective load predictions during situations such as pandemic lockdown.