Improving Energy Consumption Forecasting through Conditional Generative Adversarial Networks

Obtaining clean, high-frequency energy consumption data for households is challenging, further complicated by the presence of abnormalities in the available data. This study addresses this gap by utilizing Conditional Generative Adversarial Networks (CGANs) to generate synthetic, anomaly-free data that mirrors real consumption patterns. By incorporating time-of-day information into the CGAN model, we enhance forecasting accuracy, demonstrating a reduction in the Root Mean Square Error (RMSE) compared to traditional methods. Applied to the Electricity and Occupancy (ECO) Dataset, our approach showcases the efficacy of CGANs in addressing data quality issues in energy management. This presents a novel, cost-effective strategy for accurate energy consumption forecasting. The study compares three datasets for energy consumption forecasting: unmodified anomalies, anomalies replaced with historical mean consumption, and anomalies replaced with CGAN-generated synthetic data, with RMSE improvements observed from 255.896 to 249.336. Our findings highlight the potential of CGANs to significantly improve data quality and forecasting precision in household energy management.