Dirichlet Sampled Capacity and Loss Estimation for LV Distribution Networks With Partial Observability

Abstract

With low voltage (LV) distribution networks increasingly being re-purposed beyond their original design specifications to accommodate low carbon technologies, the ability to accurately calculate their actual spare capacity is critical. Traditionally, within the Great Britain (GB) power system, there has been limited monitoring of LV distribution networks, making this difficult. This paper proposes a method for estimating spare capacity of unmonitored LV networks using demand data from customer Smart Meters. In particular, the proposed method infers existing LV network capacity, as well as losses, across scenarios where only a limited number of customers have Smart Meters installed. Typical daily load profiles across customers with Smart Meters are learned using a Dirichlet sampled Gaussian mixture model (GMM). Learned profiles are then applied to all unmetered customers to estimate network parameters. Method accuracy is assessed by comparing estimations with simulated, fully observed, LV network models. The method is also compared to benchmark models for establishing unobserved demand profiles. Overall, results in the paper show that the proposed method outperforms benchmark models in terms of accurately assessing substation headroom, particularly in scenarios where only 10-50% of customers have Smart Meters installed.

Scenario generation of residential electricity consumption through sampling of historical data

“…Telford et al [15] assume that, for each LV feeder in the grid, a certain percentage of consumers have load time series available. Per feeder, the load time series are clustered using a Gaussian mixture model and subsequently new load scenarios are generated for unmeasured consumers of that feeder through a Dirichlet sampling procedure of the learned mixture model.…”

“…This way, we can evaluate the scenario generation directly, independent of the downstream task. Most papers look at cluster validity indices [14] to evaluate clustering or at some quantity after the power flow calculations [4,15].…”

Probabilistic load forecasting for the low voltage network: Forecast fusion and daily peaks

“…However, there is a shortage of open-source data sets that contain a satisfactory amount of nodes due to a lack of widespread monitoring. An alternative option is to generate a synthetic LV dataset from smart meter data and basic information on the local network architecture [15]. In [16] several probabilistic methods (quantile regression, KDE) are evaluated using a dataset comprising of 100 real LV feeders where there was no clear best forecaster at all feeders, however, autoregressive type models performed well, and (forecast) temperature was shown to have negligible influence on the forecast skill.…”

Probabilistic load forecasting for the low voltage network: forecast fusion and daily peaks

“…However, there is a lack of open-source available data sets which contain a satisfactory amount of nodes due to a lack of widespread monitoring. An alternative option is to generate a synthetic LV dataset from smart meter data and basic information on the local network architecture [12]. In [13] several probabilistic methods (quantile regression, KDE) are evaluated using a dataset comprising of 100 real LV feeders where there was no clear best forecaster at all feeders, however autoregressive type models performed well, and (forecast) temperature was shown to have negligible influence on the forecast skill.…”