Comparative Analysis of Deep Learning Models for Long-Term Electricity Demand Forecasting in Bangladesh Using Web-Scraped Data

As energy demand continues to rise in Bangladesh, there is a growing need for more accurate forecasting methods to improve the balance between electricity supply and consumption. Despite increased generation capacity, the country still experiences frequent disruptions due to limitations in prediction accuracy and structural inefficiencies within the power system. This research carries out an evaluative comparison of notable deep learning (DL) frameworks, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BiLSTM), and Bidirectional GRU (BiGRU), for forecasting daily peak electricity demand at both the national level and across the country's eight divisions. The dataset was compiled from the Bangladesh Power Development Board (BPDB) using an automated web scraping pipeline. All models were trained on four years of historical data and evaluated using a one-year testing set. Among the models assessed, the BiGRU architecture outperformed others, achieving the lowest testing Mean Absolute Percentage Error (MAPE) value of $4.75\%$. The BiGRU model was also employed for division-wise forecasting, effectively capturing regional demand variations. Additionally, it was applied to unseen future dates, which are not included in the dataset, where it recursively predicted energy demand one day at a time and achieved an MAPE of 7.3%, demonstrating strong generalization capability. These results signify the aptitude of deep learning-based methodologies for enabling resilient and scalable energy consumption modeling.