This thesis investigates the applications of non-parametric approaches for probabilistic demand forecasting in power distribution systems. This thesis develops two probabilistic short-term load forecasting models. We implement and evaluate two type of probabilistic forecasting methods namely: kernel density estimation and mixture density networks. In particular we are interested in the study of the features and (any) advantages of using machine learning approaches over the more traditional approaches in probabilistic demand forecasting.

This thesis gives a short-term load forecast of the residential demand with respect to the outside temperature using the probabilistic forecasting methods. The factors impacting the performance and accuracy of the forecasts are evaluated. The historical data for energy consumption generally has multiple seasonality’s associated with it. For more accurate demand forecasting, it is critical to take into account the different seasonality’s in the data and the effect of exogenous variables (temperature) while developing different models. Both the models are trained separately for yearly and seasonal datasets to study the effect of seasonality on forecasting.

Various tests are performed on the models to assess their statistical significance when compared to one another. The comparative assessment between Mixture Density Networks and Kernel Density Estimation also advances the knowledge of applying these techniques to STLF. The proposed approaches are compared with other benchmark models like ARIMA (1,0,0) model and a neural network which are also trained separately for yearly and seasonal datasets.