Repository logo

Short Term Probabilistic Load Forecasting at Local Level in Distribution Networks



Journal Title

Journal ISSN

Volume Title





Degree Level



Along with the growing inclusion of smart technologies into the electrical power grids, benefits, which can be originated form advanced metering infrastructure (AMI), have grabbed noticeable attention from distribution utilities. Since the number of meters are severely ample in practical systems, the utilities now is able to create virtual meter data by aggregating loads for distribution substations, feeders, transformers, or regions with the help of geographic information system. Such an important change brought by smart meter rollout is considered as the main factor which motivates this thesis to delve more into the load pattern modeling and forecasting at local level and find approaches which can yield to the enhanced applications in distribution networks. However, low aggregation level leads to high volatile load characteristic. In this regard, this thesis proposes a comprehensive methodology for uncertainty modeling and short-term probabilistic load forecasting (STPLF) in distribution networks. Existing methods related to uncertainty modeling and forecasting are rarely applied to local level loads and they suffer from over- or under-fitting of data when there is a misfit between the complexity of the model and the amount of data available. These models are limited to specific situations due to the great diversity of loads in distribution networks and need to be tuned every time when the load aggregation level changes. They also need a relatively large data set to support the recovery of the predictive densities. Our proposed method addresses this issue and is based on Bayesian nonparametric model which has unbounded complexity and allow the complexity to automatically grow and be inferred from the observed data. The uncertainty underlying load patterns can be endowed with any type of prior distribution and is given in a nonparametric form, i.e. a mixture model with countably infinite number of mixtures, inferred from the posterior using the Gibbs Sampling, which is a Markov Chain Monte Carlo (MCMC) technique. All effective samples from the sampling procedure along with the exogenous variables are fed to an ensemble learning machine. The final result of the probabilistic load forecasting (PLF) is averaged on the outputs of all learning models, thus reducing the model variance and enhancing the model consistency. The proposed method is tested on both a public data set and a local data set from the Saskatoon Light &Power AMI Meter Replacement Program which offers electricity consumption at a granularity of 30 minutes of more than 65,000 electricity customers including industrial, commercial and residential sectors in the city of Saskatoon, Canada.



Dirichlet process mixture model (DPMM), ensemble learning, Markov chain Monte Carlo (MCMC), probabilistic modeling, load forecasting, smart meter



Master of Science (M.Sc.)


Electrical and Computer Engineering


Electrical Engineering


Part Of