Dealing with change: Retraining strategies to improve load forecasting in individual households under Covid-19 restrictions

Abstract

Covid-19 pandemic and the restrictions imposed to control its spreading had a profound impact on several aspects of society. Studies on energy consumption show a pattern shift during the first stages of the pandemic, mainly due to the abrupt and forced change in human behaviour. For some residential consumers, such changes translated into higher energy expenses, which could be avoided through optimal energy management. This, however, depends on accurate load forecasting models whose performance tends to decrease under drastic pattern changes. This paper aims to improve forecasting techniques based on machine learning methods to assess its impact on the accuracy of day-ahead electrical load forecasting from individual residential households from the city of Barcelona, Spain, during a lockdown situation. Support Vector Machine (SVM) is the proposed algorithm for the short-term hourly forecasting of the households’ electricity consumption of the following 24 h. The algorithm’s accuracy over different pandemic periods – Lockdown, Reopening and New Normal – is evaluated using different retraining strategies based on literature findings.

Keywords

• COVID-19
• Energy demand
• Residential load
• Load forecasting
• Time series
• Support vector machine