Citation

Bushra, B. N. and Rajesh, K. and You, Ah Heng and Wong, Wai Kit and Ng, Poh Kiat (2026) A Novel Dual-horizon Forecasting of Load and Photovoltaic Generation Using a Deep Learning Framework for Smart Residential Energy Management. International Journal of Intelligent Engineering and Systems, 19 (5). pp. 775-800. ISSN 2185-3118

Text 2026053145-2.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

Smart residential energy forecasting supports the operation of smart homes by estimating future electricity demand and on-site photovoltaic (PV) generation for informed energy management decisions. Accurate multi horizon prediction is essential for real time control, day ahead scheduling and cost aware interaction with the power grid. Conventional forecasting methods relying on simplified assumptions and fixed temporal structures exhibited limited capability in modeling nonlinear consumption patterns, user driven variability, and intermittent renewable generation. Utilizing short-term and day-ahead forecasting as independent tasks further reduced applicability in integrated residential energy systems. This study presents a dual-horizon deep learning (DL) framework for simultaneous forecasting of residential load and PV generation at (t + 15 min) and (t + 1 Day) within an integrated architecture. The methodology integrates temporal convolution to extract localized patterns, recurrent neural network (RNN) and bidirectional long short-term memory (Bi-LSTM) networks to model sequential dependencies, and multi-head self attention (MHSA) to capture global temporal relationships. A gated dual-horizon embedding adaptively fuses short- and long-term contextual information to represent rapid load variations and daily consumption behavior. The model was evaluated using a Home Energy Management System (HEMS) dataset with 15-minute sampling resolution. Experimental results showed that the approach obtained a 0.1331 Root Mean Square Error (RMSE) and a 0.1321 Mean Absolute Error (MAE) for load forecasting at (t + 15 min) and an RMSE of 0.0633 with a 0.8576 coefficient of determination (R2) at (t + 1 Day). For PV generation, RMSE values of 0.1010 and 0.1030 were obtained at (t + 15 min) and (t + 1 Day), respectively, with R2 values exceeding 0.91. This study paves the way for a scalable and effective approach for multi-horizon forecasting in smart residential energy systems and supports advanced energy management and decision support applications.

Item Type:	Article
Uncontrolled Keywords:	Smart residential energy forecasting, Multi horizon prediction, Recurrent neural network, Deep learning, Bidirectional long short-term memory
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	31 Jul 2026 05:22
Last Modified:	31 Jul 2026 05:22
URII:	http://shdl.mmu.edu.my/id/eprint/16397

Downloads

Edit (login required)