This Ph.D. thesis deals with the problem of probabilistic short-term load forecastingat the low voltage level in power distribution networks. The research goal is to developa new solution that considers load variability and offers high forecasting accuracywithout excessive hardware requirements. The proposed solution is based on the applicationof statistical methods and machine (deep) learning methods for data representation(feature extraction and selection), clustering, and regression. The efficiencyof the proposed solution was verified in a case study on real smart meter data. Thecase study results confirm that the application of the proposed solution leads to highforecast accuracy and short execution time compared to related solutions.