Paper Title
A COMPARATIVE BENCHMARK OF MACHINE LEARNING AND DEEP TEMPORAL MODELS FOR URBAN LOAD FORECASTING

Abstract
The accelerated urbanization and electrification have stimulated an unprecedented increase in the electricity consumption of cities and made efficient short-term load prediction imperative to the reliable functioning of electrical grids and to intelligent economic planning. The challenge faced by high-resolution prediction is especially due to the non-linearity of demand behaviour as intrinsically problematic, the unpredictability of weather, and the changing utilization patterns. This paper proposes an analytical reference framework of the ultra-short-term power demand forecasting of cities, based on a multi-year dataset of the city of Delhi which is measured at every five minutes and enhanced with meteorological measurements. It has designed coherent preprocessing and feature-engineering pipeline to include temporal cycle, autoregressive memory and environmental context. Three or three symbolic modelling paradigms are thoroughly compared: a histogram-based gradient-boosting regress or of structured learning, a long-short-term memory network of sequential modelling and a temporal convolution network of the extraction of long-range dependencies. Empirical results show that gradient-boosting consists of the minimization of the forecasting error, with deep temporal models coming into the same competition but requiring more computational resources. Its results provide practical recommendations to utilities, which show that highly designed machine learned solutions can be very accurate with less operational complexity to be deployed in real-time. Keywords - Data analytics, deep learning, electricity demand forecasting, gradient boosting, load prediction, temporal convolution networks, time-series analysis, urban energy.