Machine Learning For Fast Short-Term Energy Load Forecasting

Abstract

Improved energy usage data from smart meters offers a unique opportunity to apply advanced analytics that can dramatically improve load forecasting. Utility companies, policy makers, and consumers benefit with better integration of renewables and overall energy management in the IoT digital age. Accurate short-term energy forecasting is essential to improving energy efficiency, reducing blackouts, and enabling smart grid control. In this work-in-progress (WIP) paper, we use individual residential load data to perform customer segmentation based on energy profiles, introduce a unique data segmentation and feature extraction technique based on inherent load signal periodicities, and use deep learning to perform fast and accurate short-term forecasting. Partnering with Prime Solutions Group, a veteran-owned company based in Arizona, we found that we could obtain up to a 12% improvement in hourly one-day forecasting using our custom data segmentation and feature extraction techniques with neural network methods.