Daily Peak Load Forecasting Using Smart Meter Data

There is a need for greater visibility, understanding, and flexibility of electricity demand to support the transition to low-carbon energy. The smart meter roll out in Great Britain presents new opportunities to address these three challenges for all energy consumers. For instance, dynamic pricing allow households to consume low-cost energy more during periods of high renewable output and provide flexibility as an alternative to expensive network reinforcements. Short-term forecasts of individual electricity demand will be essential to provide coordinate flexibility efficiently, especially with proliferation of electric vehicles, heating, and batteries [1].

Forecasting at low voltage levels can be challenging, and different to the conventional approach of demand forecasting at the transmission level. As electricity demand is aggregated, group behaviors become clear, as shown by the emergence of a smooth diurnal pattern in Figure 1, which tend to only change slowly and are therefore relatively predictable. Disaggregate demand is much more changeable and influenced by unknown behaviors and processes. This requires a new approach to forecasting, which should be developed with the end-use in mind.

Figure 1. Average half-hourly demand for a collection of smart meters (SM), with the number of SMs indicated at the top of each panel, during a week in November 2013. As the number of smart meters increases the time series becomes smoother and more predictable.

Peak demand is typically the limiting factor in the capacity of distribution networks, set by the maximum power a cable of transformer can handle. Therefore, day-ahead forecasts of the peak demand could be useful from both utilities perspective (e.g. in setting dynamic prices), and the consumers perspective (e.g. for scheduling battery or EV charging).

Another aspect to consider is that electricity demand is hierarchical from the meter, feeder, secondary substation, primary substation level, and above. Respecting this reality in our forecasting models can help improve accuracy and enable more coordinated decisions at different levels of the network. A plot of the daily peak demand at the primary substation level is shown on Figure 2 for a hypothetical hierarchy constructed from the bottom-up, using smart meter data.

Figure 2. Daily peak demand for a hypothetical (& small!) primary substation constructed from aggregate smart meter data. At this level of aggregation there is seasonal, weekday, and temperature dependencies which help inform forecasting models.

Of course peak demand intensity is only one half of the problem; considering the time-of-peak is also very relevant. There is significant diversity in the time-of-peak characteristics at the smart meter level, driven by individual behaviours, as shown in Figure 3. Therefore, the goal of our research is to forecast bivariate quantities of daily peak intensity and timing at these different voltage levels, with a view to create coherent probabilistic forecasts of these quantities.

Figure 3. Histogram of the time-of-peak in a hypothetical hierarchy at the smart meter (sm), feeder (fdr), and substation (ss,ps) levels. At the SM level of aggregation the time-of-peak is spread over the full day, and as the aggregation level increases peaks become more concentrated in the evening.

Finally, there is no one-size-fits all approach to forecasting disaggregate demand, given the diversity of behaviours. To illustrate, Figure 4 shows the autocorrelation and partial autocorrelation function of the daily peak demand intensity for 4 smart meters. This suggests that using auto-regressive type models could be successful at some, and not at others. Thanks for reading!

Figure 4. Autocorrelation (acf) and partial autocorrelation (pacf) plots of daily peak demand at 5 smart meters. There is a diversity of characteristics of the time series suggesting that different model formulations will work better at different meters, e.g. autoregressive type models could perform well at N0655, but not N0304. Clustering smart meters into similar groups will help address this problem.

[1] – https://arxiv.org/pdf/2005.10092.pdf

data source: http://doi.org/10.5255/UKDA-SN-7857-2

AMIDiNe Begins

The programme of research that constitutes AMIDiNe will address the increasingly problematic management of electrical load on distribution networks as the UK transitions to a low carbon energy system. Traditionally, distribution networks had no observability and power flowed from large generation plant to be consumed by customers in this ‘last mile’. Now, and even more so in future, those customers are generators themselves and the large generators that once supplied them have been supplanted by intermittent renewables. This scenario has left the GB energy system in position where it is servicing smaller demands at a regional or national level but faces abrupt changes in the face of weather and group changes in load behaviour, therefore it needs to be more informed on the behaviour of distribution networks. Increasing availability of Smart Meter data through the Data Communications Company has the potential to address this but only when placed within the context of analytical and physical models of the wider power system.

The UK government’s initiative to roll out Smart Meters across the UK by 2020 has the potential to illuminate the true nature of electricity demand at the distribution and below levels which could be used to inform network operation and planning. Current research broadly involving Smart Metering focuses on speculative developments of future energy delivery networks and energy management strategies. Whether the objective is to provide customer analytics or automate domestic load control, the primary issue lies with understanding then acting on these data streams. Challenges that are presented by customer meter advance data include forecasting and prediction of consumption, classification or segmentation by customer behaviour group, disambiguating deferrable from non-deferrable loads and identifying changes in end use behaviour. Transferring this to an operational scenario will require understanding of where and how computing resource is placed and specifically which functionality we are asking of these resources, the interchange formats between substation automation data and premises level, the operational forms of models with which the intelligent grid is informed and the predicates on which they act.

Moving from a distribution network with enhanced visibility to augmenting an already ‘smart’ transmission system will need understanding of how lower resolution and possibly incomplete representations of the distribution network(s) can inform more efficient operation and planning for the transmission network in terms of control and generation capacity within the context of their existing models. Improving various distribution network functions such as distribution system state estimation, condition monitoring and service restoration is envisaged to utilise analytics to extrapolate from the current frequency of data, building on successful Big Data techniques already used in other domains. This extrapolation would require significant scaling of the analytics used to make Distribution System State Estimation and Dynamic Optimal Power Flow techniques useful at a full network scale and would be able to support the introduction of automation. The overall aim is that strategic investment decisions for network infrastructure components can be made on the back of improved information availability. These decisions could be deferred or brought forward in accordance with perceived threats to resilience posed by overloaded legacy plant in rural communities or in highly urbanised environments; similarly, operational challenges presented by renewable penetrations could be re-assessed according to the true nature of demand and export profiles and their relation to network voltage and emergent protection configuration constraints.