Parametric Time-series Modelling of London Smart Meter Data for Short-term Demand Forecasting

Abstract

Electricity being one of the most important components behind economic growth in 21st century, accurate electricity demand forecast became essential. Now with the deployment of smart meters that are capable of providing half-hour energy usage data comes new opportunities for short-term demand forecasting. In this research two statistical timeseries models known as the seasonal auto-regressive integrated moving average (SARIMA) and with exogenous inputs (SARIMAX) are employed to study half-hourly energy demand forecast and daily peak forecast capability over a week at half-hourly interval. The models are tuned and tested on a half-hourly aggregate level data and individual meters data extracted from London smart-meter dataset. The models are also cross validated over different seasons to evaluate model robustness over different training data size and forecasting under different temperature conditions. The SARIMA model performed better at consistently forecasting daily-demand peaks, while the SARIMAX was overall more accurate as compared to the SARIMA at more computational cost. This is because of the exogenous temperature variable used in SARIMAX which explains some of the demand profile volatility due to temperature changes. This also resulted in a better fit for the SARIMAX model. The models tested in this paper can accurately forecast energy-demand at half-hour intervals and daily-peaks for a week-ahead forecast at a regional demand profile over different seasonal condition.