Journal of Innovations in Social and Applied Sciences

Double-blind peer review · Immediate open access
Volume 1 (2025), Issue 2

Deep Learning-Based Energy Demand Forecasting for Sustainable Smart Cities

Amr E. Elshora Department of Computer Science and Information, Higher Institute of Management in Kafrelsheikh – Ministry of Higher Education
Open Access — CC-BY-4.0
Download PDF Read Online

Keywords: energy demand forecasting; smart cities; deep learning; LSTM; smart grid; household electricity; time-series forecasting
Abstract

Reliable short-term electricity forecasts have a direct role in smart-city operations, including storage scheduling, peak control, and demand response. In this paper, we examine a direct multi-horizon forecasting strategy using the public Individual Household Electric Power Consumption dataset. The original 2,075,259 minute-level measurements were aggregated to 34,589 hourly observations. For each forecasting sample, the model uses 72 hours of seven electrical variables to predict Global Active Power for the next 24 hours. The chronological order of the series was preserved by assigning 70% of the data to training, 15% to validation, and 15% to testing, which produced 24,117, 5,165, and 5,166 windows, respectively. The comparison includes a persistence baseline, a standard LSTM, a GRU, and a Compact Stacked LSTM. On the held-out test period, the GRU produced the lowest error, with an RMSE of 0.6181 kW and an MAE of 0.4687 kW. The standard LSTM followed closely with an RMSE of 0.6262 kW, while the stacked LSTM recorded 0.6509 kW. Persistence performed considerably worse, with an RMSE of 0.9177 kW. These results support recurrent modelling for this forecasting task, but they also show that simply adding recurrent depth does not necessarily improve household-demand forecasts.

How to Cite
Amr E. Elshora (2025). Deep Learning-Based Energy Demand Forecasting for Sustainable Smart Cities. Journal of Innovations in Social and Applied Sciences (JISAS), 1(2), 48-61.