No abstract

Detailed large scale simulations require a lot of data. Residential electrical load profiles are well protected by privacy laws. Representative residential electrical load generators get around the privacy problem and allow for Monte Carlo simulations. A top-down model of the residential electrical load, based on a dataset of over 1300 load profiles, is presented in this paper. The load profiles are clustered by a Mixed Model to group similar ones. Within the group, a behaviour model is constructed with a Markov model. The states of the Markov models are based on the probability distribution of the electrical power. A second Markov model is created to randomize the behaviour. A load profile is created by first performing a random-walking the Markov models to get a sequence of states. The inverse of the probability distribution of the electrical power is used to translate the resulting states into electrical power.

Two problems are tackled in this paper: determining the active demand reduction potential of wet appliances and making time series estimates from project data. The former is an application of the latter. Household groups representative to the average population are defined by applying Expectation Maximization clustering to a representative measurement set (n = 1363). Attitudes towards active demand are found by conducting a survey (n = 418). Project data (n = 58) containing wet appliance measurements are scaled up by adapting the clustering algorithm, spreading the electricity demand of the wet appliances over the clusters. The potential for active demand reduction with wet appliances is 4% of the total residential power demand, assuming that 29% of the households take part. The potential is in the order of magnitude of the power reserves, but does not fulfill availability and response time requirements.

Smart grid pilot projects require a representative subset of the total population to draw relevant conclusions from test results. However, customers willing to participate in such projects are not always representative to the whole population. Standard random sampling gives some problems because not all results can be scaled. Defining sub-populations or strata to random samples from is theoretically sound, but the definition of sub-populations is quite expensive. The paper presents a customer sampling technique based on quota. The domains for the quota are defined by machine learning algorithms and the quota themselves are based on realistic data. Sampling is done by an optimization algorithm, which eliminates the common 'human error'-factor in quota sampling. The approach is a cost efficient and convenient way of sampling that is able to balance the representativeness of the electricity consumption patterns for the population against sampling accuracy. The method has been applied and validated on a large customer data set.

No abstract