Short-term load analysis and forecasting using stochastic approach considering pandemic effects

The COVID-19 pandemic and its containment have changed the pattern of electricity load. Hence, accurate forecasting of load for shorter interval of time during pandemic has become extremely challenging. Application of machine learning (ML) and deep learning (DL) techniques has improved accuracy to forecast load. This study compares ML and DL algorithms for short-term load forecasting. This research introduces a novel 'COVID Operator' for load forecasting. The operator is characterized by three distinct levels, namely 0, 1, and 2. The '0' denotes pre-COVID, '1' signifies during, and '2' designates post-COVID. Unlike previous studies that just looked at the beginning of the pandemic, this study examines electricity load data from March 2017 to January 2021. Outliers of data are removed using interquartile range (IQR) method. After outliers are removed, data is trained using five different technique-based ML algorithms viz Gaussian Process Regression (GPR), Ensemble, support vector machine (SVM), Light Gradient-Boosting Machine (LightGBM) Categorical Boosting (CatBoost) and DL algorithm viz Deep Neural Network (DNN). After comparing all methods, GPR is the best method due to a good trade-off between training time and root mean square error (RMSE), and its ability to sensitivity to load changes. GPR is further hyperparameter tuned with Grid search, Random Search, and Bayesian Optimisation with an objective of reducing mean square error (MSE). Due to hyper-parameter tuning, the MSE of GPR is decreased by 10.3% with Grid Search being the optimal choice. Moreover, in this study, February 2021 data set is considered as 'Subject Out' set, and tested using developed algorithms. It is found that GPR is having less error in terms of RMSE as 64.38 MW and similar results for other performance parameters.