Day-Ahead Building Power Demand Forecasting in Smart Grids

In this dissertation, we propose a novel day-ahead load forecasting method that can be applied without manual setup on any building and is more accurate than currently existing methods for predicting low-voltage loads. Day-ahead predictions allow a smart grid to mitigate the volatility of decentralized renewable generators locally, by using demand flexibilities of the buildings located in the area. Historically, power system operators forecast low-voltage demand for the upcoming day using standard load profiles. While this basic method is effective for large consumer aggregations, it lacks accuracy when applied on smaller loads and the flexibility to consider modern energy equipment in the buildings. More advanced forecasting methods that exist for the high-voltage level, rely on manual fine-tuning and can be used only in singular cases. Our aim is to develop a method that can replace standard load profiles for predicting low-voltage loads on a wide scale -- a method that can be applied on numerous individual buildings of different size and type without any explicit knowledge of them.

We formulate the wide-scale day-ahead load forecasting problem in low-voltage domain studying various loads and their characteristics. Considering time-series nature of the data and its nonstationarity, we combine nonparametric functional data analysis with the theory of statistical learning to introduce a univariate autoregressive functional neighbor model with corresponding forecasting algorithm. Additionally, we present an extension that allows to consider exogenous variables which can affect the consumption of a given building. We evaluate the model on an extensive, publicly available dataset of loads and use inferential statistics comparing our model to numerous references.

The main result of this work is a load forecaster that can be universally applied in a distribution grid using historic load measurements and, optionally, further inputs. Statistical analysis shows that our model can be expected to be significantly more accurate than standard load profiles and more sophisticated approaches based on classical time series analysis and machine learning. Even for the largest loads, our method can be expected to be at least 39% more accurate than standard load profiles that were designed to predict larger aggregations of end-consumers. Therefore, given mass adoption of smart meters, the proposed functional neighbor model can replace standard load profiles that were used in power systems since their inception. Improved accuracy and flexibility of the proposed method facilitates various smart grid applications that can increase the efficiency of the existing distribution system infrastructure and aid accommodating renewable energy generators.