Long-Term Electrical Load Forecasting: A Case Study for the Chattogram, Bangladesh

Long-term future transmission and generation capacity expansion underscore the importance of long-term electricity load forecasting. This paper conducts the annual electricity peak demand prediction using machine learning (ML) methods such as linear regression (LR), artificial neural network (ANN), XGBoost, recurrent neural network (RNN), and long-short-term memory (LSTM). Univariate lag features are used to train the models. The historical load data from 1971 to 2021 is split into the ratio of 8:2 for training and testing datasets respectively. Models’ robustness is examined to add Gaussian white noise with various intensities. Explainable AI using Shapley Additive exPlanations (SHAP) is employed to interpret the block-box character of the best ML technique. Linear Regression model yields the most accurate prediction results compared to XG Boost, ANN, RNN, and LSTM. Models’ performance accuracy is examined based on RMSE (root mean square), MAP (mean absolute percentage error), MAE (mean absolute error), and R-squared. The prediction results are assessed with an economic model such as the GDP-elasticity method. The experimental results show that the electricity peak demand will grow from 2979 MW in 2024 to 10796 MW in 2041. The linear regression method estimates the annual average electricity demand growth rate is 7.29% compared to the 7.68% economic method. The linear regression is the best model for a long-term forecasting compared to XG Boost and deep learning models.