Distribution Network Demand and Its Uncertainty

Abstract

This chapter presents some advanced tools for low voltage (LV) network demand simulation. Such methods will be required to help distribution network operators (DNOs) cope with the increased uptake of low carbon technologies and localised sources of generation. This will enable DNOs to manage the current network, simulate the effect of various scenarios and run load flow analysis. In order to implement such analysis requires high resolution smart meter data for the various customers connected to the network. However, only small amounts of individual smart meter data will be available and such data could be expensive. In likelihood, smart meter data is only going to be freely available at the aggregate level. Hence, in general, to implement LV network tools, customer loads will need to be simulated based on the assumption of limited amounts of monitored data. In addition, due to the high volatility of LV electric distribution networks, demand uncertainty must also be captured within a simulation tool. In this chapter, a number of methods are described for simulating demand on low voltage feeders which rely only on relatively small samples of smart meter data and monitoring. Firstly, a method called ‘buddying’ is described for assigning realistic profiles to unmonitored customers by buddying them to a customer who is monitored. Secondly, a number of methods are presented for capturing the uncertainty on the network. Finally the uncertainty models are incorporated into the buddying method and implemented in a load flow analysis tool on a number of real feeders. Both the buddying and the uncertainty estimation are presented for two different cases based on whether LV substation monitoring is present or not. This illustrates the different impacts of monitoring availability on the modelling tools. This chapter demonstrates the presented methods on a large range of real LV feeders.

Appendix

3.1.1 Customer Groups

The customers are split into groups according to their profile class, council tax band and the presence of photovoltaic equipment. There are eight generic Elexon profile classes representative of large populations of similar customers. Two classes correspond to domestic customers and distinguish between two tariffs, “Standard” and “Economy 7”. The latter provides cheaper rates overnight at the expense of increased day-time charges. Council Tax is a local taxation system used in Great Britain essentially based on property value. Each property is assigned one of eight bands (A–H) in ascending property value.

The grouping used for residential customers is shown in Table 3.1. The commercial customers are not listed, but essentially, a separate profile is used for broad classes of customers (e.g., school, hospital, church, etc.). More details on commercial can be found in Sect. 3.2.5.

Table 3.1 Table showing grouping assignment for UK residential customers based on profile class, council tax band and whether the property has known photovoltaics

Note that the main purpose for the grouping is to reduce the computational cost of the buddying algorithms. However, many other characteristics could be used to group the customers, such as MOSAIC socio-demographics classifications, if profile class and/or council tax band are not available. For other countries similar socio-demographic/asset based groupings are also valid.

3.1.2 Genetic Algorithm

The genetic algorithm mimics the process of natural selection and proceeds by creating updates of several collections of monitored customers according to how well they score according to the fitness function (3.4). The basic steps of the genetic algorithm are outlined below, which are also summarised in Fig. 3.14.

Fig. 3.14

A graphical presentation of the genetic algorithm

• 1. Initialise the buddy. Create G genomes each consisting of M randomly selected profiles from P (see Sect. 3.2) for a training period of H half-hours for each customer $c_{j} , j = 1, \ldots , M$. The selection of the buddies is only restricted so that the buddies belong to the same group as customer $c_{j}$ (see Sect. “Customer Groups”).

• 2. The fitness of each genome is evaluated using the fitness function (3.4).

• 3. Select the best-scoring (fittest) genomes.

• 4. To create each of the G next generation genomes, two of the current best G′ < G genomes are randomly selected for crossver. Common profiles are retained while the remaining profiles are selected randomly from one or the other genome.

• 5. The new genomes are mutated by replacing each profile with a probability p with a new profile (from the same group).

• 6. Repeat steps 1–3 for 100 generations.

The probability of mutation is free to chose, but is initially set to p = 0.1 and slowly decreases as the algorithm progresses. A mutation rate too low and the genomes may lose variation, too high and good solutions may be removed from the population. For step 3, after 40 iterations the genomes are reset, whilst retaining the best genome, to reduce the chances of finding a local minimum.