Konstantinos Theodorakos

Long-term individual household forecasting is useful in various applications, e.g., to determine customers’ advance payments. However, the literature on this type of forecasting is limited; existing methods either focus on short-term predictions for individual households, or long-term prediction at an aggregated level (e.g., neighborhood). To fill this gap, we present a method that predicts the monthly consumption of individual households over the next year, given only a few months of consumption data during the current year. Utility companies can exploit this method to predict the consumption of any customer for the next year even with incomplete data. The method consists of three steps: clustering the data using ${k}$ -means, prediction using an ensemble of forecasts based on the historical median distribution among similar households, and smoothing the predictions to remove weather-dependent patterns. The method is highly accurate as it finished third in the IEEE-CIS competition (and ranks first when leveraging insights from another team), focused on forecasting long-term household consumption with incomplete data. It is also very scalable thanks to its low computational complexity and weak data requirements: the method only requires a few months of historical data and no household-specific or weather information.

Air pollution causes around 400,000 premature deaths per year in Europe due to Particulate Matter, nitrogen oxides and ground-level ozone pollutants. Multiple-Input Multiple-Output Nonlinear Auto-Regressive eXogenous Deep Neural Networks are frequently used to predict a day before, air-quality pollution incidents, at a country-scale. With complexity and data sizes increasing, finding performant models becomes harder. We propose island transpeciation to optimize hyperparameters and architectures. Unlike using a single optimizer, island transpeciation combines results from multiple optimizers, to consistently provide excellent performance. Moreover, we show that island transpeciation outperforms random model search and other previous modelling efforts. Island transpeciation is a Neural Architecture Search that uses co-evolution (genes), to combine (transpeciation) populations of incompatible optimizers (species) organized in island formations. In island transpeciation, architecture search is parallelized and utilizes a distributed pool of hardware resources. We have successfully used these techniques to predict next-day ozone concentrations across the Belgian territory.

Day-ahead energy forecasting systems struggle to provide accurate demand predictions due to pandemic mitigation measures. Decomposition-Residuals Deep Neural Networks (DR-DNN) are hybrid point-forecasting models that can provide more accurate electricity demand predictions than single models within the COVID-19 era. DR-DNN is a novel two-layer hybrid architecture with: a decomposition and a nonlinear layer. Based on statistical tests, decomposition applies robust signal extraction and filtering of input data into: trend, seasonal and residuals signals. Utilizing calendar information, temporal signals are added: sinusoidal day/night cycles, weekend/weekday, etc. The nonlinear layer learns unknown complex patterns from all those signals, with the usage of well-established deep neural networks. DR-DNN outperformed baselines and state-of-the-art deep neural networks on next-day electricity forecasts within the COVID-19 era (from September 2020 to February 2021), both with fixed and Bayesian optimized hyperparameters. Additionally, model interpretability is improved, by indicating which endogenous or exogenous inputs contribute the most to specific hour-ahead forecasts. Residual signals are very important on the first hour ahead, whereas seasonal patterns on the 24th. Some calendar features also ranked high: whether it is day or night, weekend or weekday and the hour of the day. Temperature was the most important exogenous factor.