Electricity Demand and Population Dynamics Prediction from Mobile Phone Metadata

Abstract

Energy efficiency is a key challenge for building modern sustainable societies. World’s energy consumption is expected to grow annually by 1.6 %, increasing pressure for utilities and governments to fulfill demand and raising significant challenges in generation, distribution, and storage of electricity. In this context, accurate predictions and understanding of population dynamics and their relation to electricity demand dynamics is of high relevance.

We introduce a simple machine learning (ML) method for day-ahead predictions of hourly energy consumption, based on population and electricity demand dynamics. We use anonymized mobile phone records (CDRs) and historical energy records from a small European country. CDRs are large-scale data that is collected passively and on a regular basis by mobile phone carriers, including time and location of calls and text messages, as well as phones’ countries of origin. We show that simple support vector machine (SVM) autoregressive models are capable of baseline energy demand predictions with accuracies below 3 % percentage error and active population predictions below 10 % percentage error. Moreover, we show that population dynamics from mobile phone records contain information additional to that of electricity demand records, which can be exploited to improve prediction performance. Finally, we illustrate how the joint analysis of population and electricity dynamics elicits insights into the relation between population and electricity demand segments, allowing for potential demand management interventions and policies beyond reactive supply-side operations.