Time Series Forecasting, Applications on Low Voltage Grid Data

In an era marked by the pressing need to address climate change, innovative solutions in energy management are becoming increasingly crucial. In this thesis, three main contributions to the field of energy forecasting are presented. The first contribution is a novel method for long-term forecasting of individual household electricity consumption, demonstrating high accuracy and scalability. The second introduces a global probabilistic approach for short-term forecasting of individual households consumption, particularly robust to changing patterns. The third establishes an innovative pipeline for optimizing energy efficiency in office buildings through automated plug on and off scheduling. Each of these contributions underscores the pivotal role of advanced forecasting in energy management. But how exactly does energy time series forecasting intertwine with the broader goals of the energy transition?

Global efforts intensify to address the climate crisis, and this is resonating through headlines, policy agendas, and conversations at all levels of society. To meet the latest COP agreements and related European green deal towards a sustainable future, we must speed up electrification, increase investments in renewable energy sources and improve energy efficiency. Accurate energy forecasting then emerges as indispensable. For example, as the transition from traditional vehicles to electric cars accelerates, the demand for home charging and public charging infrastructures intensifies. This not only increases overall electricity consumption but also poses the risk of grid congestion, especially during peak times when many return home to charge simultaneously. To efficiently manage and mitigate such challenges, precise forecasting of household electricity demand becomes imperative.

Energy forecasting literature has traditionally focused on long-term or coarse spatially aggregated predictions. These predictions typically represent national-level or production-oriented estimates. The introduction of smart meters and smart plugs has made new data available. This allows for more detailed forecasting, at the individual household level or at the device level. Yet, inherent challenges accompany this shift. First, even though data is becoming more accessible, obtaining datasets remains challenging, largely because of privacy concerns. Additionally, it is crucial to have representative datasets that capture current trends, like the rise in heat pump usage or the increase in remote work due to the COVID situation. Secondly, maintaining high data quality is crucial. Any compromise can significantly degrade predictions. Thirdly, the stochastic patterns in individual household energy use need to be specifically addressed. The spikes in consumption are primarily influenced by human behavior, like when residents return home and begin using appliances. Lastly, ensuring scalability is essential, especially when forecasts are to be applied across thousands of households, for example. 

Three principal research components form this thesis. The first tackles the so far unaddressed problem of long-term individual household forecasting. The proposed method for predicting both monthly and yearly electrical consumption of individual households one year ahead was developed applying pre-processing, data augmentation, clustering, and ensemble learning to produce precise forecasts. The method is able to predict an entire year ahead, even with limited historical data, and requires no specific household attributes. The method is scalable and accurate, as demonstrated by its top-ranking in the IEEE-CIS competition. This method can be applied to ensure accurate billing capabilities for suppliers.

The second method, producing short-term forecasts of individual household consumption, tackles the double peak penalty effect. This double penalty effect arises when point forecasting methods accurately predict a peak but with a slight timing error, incurring penalties both when the peak is unpredicted and when it is wrongly anticipated. However, predicting a peak with a slight delay is more valuable than missing it entirely. Employing probabilistic forecasts effectively mitigates this issue. A novel global probabilistic method is introduced, which requires only a week of historical data from the concerned household to make accurate one-day-ahead predictions, by leveraging historical data from other households in the dataset. Its adaptability and robustness has been assessed through evaluation across three different datasets and benchmarked with several state-of-the-art methods. The method, unlike all other benchmarks, adapts quickly when large changes in patterns occur, such as a new family moving in or a holiday period. 

The third component establishes a comprehensive pipeline to reduce office building energy consumption. This involves automatically determining if a plug is idle or active, forecasting its usage, and then scheduling it on or off. With electricity savings of the monitored plugs reaching up to 20\%, this method showcases the potential for demand flexibility and optimizing building energy efficiency. The automatic detection of idle versus active mode, using an ensemble of unsupervised clustering, fills a void in the literature. Existing approaches identify this idle/active threshold either by visual assessment or through hands-on experiments. The data, sourced from the University of California, San Diego, consists of over 150 individual plug consumption profiles, over more than a year. It is made openly accessible to the community, and is the first of its kind. 

A significant overarching contribution of this research, is its emphasis on transparency and reproducibility. By providing access to the developed code, introducing a new dataset, and always utilizing open datasets, the study ensures that both the scientific community and industry professionals can derive value from it. This open approach facilitates rigorous comparisons in subsequent research and fosters a collaborative environment for further advancements in the field.