A Deep Learning Approach to Forecast Electricity Demand in Sylhet of Bangladesh

Electricity demand forecasting is essential for efficient resource allocation, infrastructure planning, and sustainable energy management. However, achieving a precise prediction of electricity demand is difficult due to noise and randomness. This paper presents a comparative analysis of deep learning and machine learning models for forecasting electricity demand in Sylhet, a rapidly developing urban area in Bangladesh. We applied Time Series Dense Encoder (TiDE), which has not been used in this field before, to achieve higher accuracy in electricity demand forecasting. To assess the performance of the proposed approach, we compared it with three different machine learning models: linear regression, XGBoost, and an ensemble of linear regression and XGBoost. Each model’s accuracy was evaluated using MAPE, RMSE, MAE, and the coefficient of determination $\left({r}^{2}\right)$. Based on the research outcomes, the TiDE model had MAPE values of $1.139 \%, \mathbf{1. 6 7 2 \%}, 2.887 \%$; RMSE values of 0.0081, $0.0139,0.0230;$ MAE values of $0.0065,0.00972,0.0162$; and Coefficient of Determination values of $0.91,0.814,0.717$ for the three forecast horizons in these order-days seven, fifteen and thirty.