The Short-term Electricity Consumption Forecast Competition Under COVID-19 Lockdown Conditions

Imagine you are responsible for the day-ahead electricity consumption forecast, and from tomorrow, your country is quarantined — a new reality is coming. Do I need to change my forecast approach? How do I “explain” to the model that tomorrow is the first day of a new reality? What should I do if the quarantine is prolonged? This paper presents the author’s experience during a short-term electricity consumption forecast competition in Russia, the start of which coincided with the beginning of the COVID-19 lockdown (30 March – 11 May 2020). The forecast service developed by the author won the competition. The study describes: The forecast service architecture (Python-based, Docker containerized), including five models: Seasonality (S), Temperature (T), Holiday (H), Autoregression (A), and a Neural Network (N). Regress and Neural Network model stacks and their assembly for improved accuracy. Numerical results reveals that forecast errors during the lockdown were 1.5-2 times higher than under normal conditions, with specific challenges during the initial lockdown period, central heating switch-off, and the May holidays. The paper concludes with actionable answers to the three opening questions, recommending k-fold training for neural networks, treating new reality stages as separate holiday groups, and continuously monitoring blind network performance.