Residential electricity load forecasting using deep learning tech

• Main Author: Zhu, Nianyao
• Other Authors: Xu Yan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/182955
• Institution: Nanyang Technological University

In this work, we adopt the multi-scale deep learning framework TimesNet, which embeds attention mechanisms to capture both short and long-term temporal dependencies of residential electricity demand. By fusing historical consumption data together with exogenous factors like temperature, humidity, and calendar variables, the proposed approach models complex and evolving load patterns effectively. Experimental results over real datasets show that the proposed approach beats baseline approaches, such as LSTM and RNN, in terms of lower forecast error, ensuring superior performances also in regimes affected by high volatility and anomalous demand events. These results underpin the value of TimesNet to assist utilities with optimized operational decisions by shaving peak-loads and proactively informing energy management policies. Real-time deployment, interpretability analysis, and integration with spatially aware models are some of the coordinated regional forecasting issues that will be addressed in future works.