Bachelor's thesis Julian Vossen

 

Probabilistic forecasting of household electrical load using artificial neural networks

For optimizing the usage of electricity, home energy management systems require accurate forecasts
of the electrical load on a single-household level. As this load is subject to a high volatility and
unpredictable human behavior, deterministic point-forecasting methods fail to provide accurate
predictions. Towards overcoming this challenge, probabilistic forecasting methods are proposed
in order to incorporate the uncertainty into the forecasts. In this thesis, we use Mixture Density
Networks and Softmax Distribution Networks to predict the probability density over the electricity
consumption one hour into the future. The models are evaluated on three smart meter datasets,
the Smart*, a UCI and the UK-DALE dataset. Thereby, scoring functions such as the Continuous
Ranked Probability Score are used to evaluate the predictive densities while the point-forecasting
performance is measured in terms of the mean absolute error. The main finding is that both neural
network models generate reliable density forecasts which significantly outperform an unconditional
benchmarking model. When evaluating different model input configurations, conditioning models on lagged electrical load led to drastic performance gains. However, the models did not
significantly benefit fromincreasing the length of the lagged input window, whereas decreasing the
dataset granularity improved the forecasts. The results of this thesis pave the way to investigating
if the availability of reliable probabilistic forecasts can result in a more effective optimization of
household electricity consumption.