Data-driven thinking in demand response

The legacy grids are being renovated to smart grids using various flavours of sensing, communication, control and internet technologies. Demand response (DR) is one of the most successfully implemented mechanisms on the smart grid platform as it exists today. This has enabled the active participation of consumers in providing flexibility through grid balancing services (once the preserve of centralised generators) – thereby reducing costs and carbon.

A major share of demand side flexibility is offered by large industrial and commercial building consumers using their energy assets such as curtailable loads, standby generators and storage systems. The availability of these energy assets for DR is dependent on the building’s energy consumption which in turn is dictated by the building’s operational schedules, occupant behaviour and ambient weather, among other factors.

Let us explore a few examples to understand this:

• The share of HVAC (heating, ventilation and air-conditioning) load in a typical office building varies with the time of day and is generally higher during the working hours. As a result, higher HVAC load curtailment is available during these hours. However, what would be the consequence of switching off that electric space heating system for hours on end on a cold winter day in a heroic attempt to save the grid? Freezing and disgruntled occupants of course! Hence it is important to account for factors that influence the building’s energy consumption, in order to estimate the availability of curtailable loads.
• Some buildings may deliver DR by aligning load test runs of a standby generator to the market’s need for electricity. Where there is no export licence, the curtailment capacity provided by the generator to the grid depends on the building load backed up at that time. Being able to predict the building load for different time periods helps commit the DR capacity from standby generators more accurately.
• Another area to consider is battery storage (technology that some say has the potential to save the planet). Imagine a manufacturing plant with a large-volume behind-the-meter battery delivering the evening peak management (red zone or triad) DR to the grid. Prior to delivery, ample energy is required in the battery to back up the predicted building load for that period.  If DR events start to draw that reserve down towards unacceptable levels, consumption must be increased for a time to recover the charge, affecting net building load.

These examples highlight the importance of building load prediction in delivering DR from different energy assets. For a building manager, commitment of DR availability with good accuracy results in higher incentives and lower penalties. For the system operator, the improved accuracy in DR availability from the consumer-side helps in better system planning. This being the kind of situation that Flexitricity as an aggregator strives to create, we asked ourselves – how do we achieve this?  

Theory-based physical models developed using tools such as EnergyPlus help predict building energy consumption with a high level of accuracy. However, they demand detailed information on the building geometry and thermal parameters that are often inaccessible and can change from time to time (the thermal inertia of a supermarket changes with stock type and volume, for example). Considering the large number of sites involved in DR, such a theory-based approach towards building load modelling is cumbersome and impractical. That’s why data-driven alternative approach is preferable in many cases.

Smart grids provide the opportunity to access ‘big data’ from the electricity system in near real-time. These include consumption data from smart meters, weather data and process measurements. Progress in the field of machine learning, availability of off-the-shelf tools and advancement in computational capabilities help us make use of this data.

Data-driven models are tuned by the data themselves, enabling custom model development for the individual sites and energy assets located there. Once deployed, any internal changes to the building (for example addition of a new load, refurbishment, etc.) will be reflected in the incoming smart meter data, to which the data-driven model will adapt without any human intervention – this is the beauty of artificial intelligence (AI). Another enticing aspect is that the modelling methodologies are replicable on large number of buildings without incurring additional costs. 

Data-driven models could be deployed in different DR operational contexts. For example, a week-ahead prediction model developed for an office building may be deployed for frequency response programmes which mandate week-ahead capacity commitments. Data-driven models may also be used for baselining purposes to evaluate the actual delivered response against what was committed. Other applications include peak demand and price prediction. Setting up a pipeline of data collection-processing and model development-deployment could motivate data-informed decision making in many other DR contexts. Having foreseen this, Flexitricity is already developing intelligent machines that work hand-in-hand with our operational team in order to maximise profits for our Energy Partners and benefit the grid.    

Going beyond DR, data-driven thinking is essential in the wider context of electricity system operation. Some of the inspiring AI applications in this sector are: reduction in energy consumption, smart electricity theft detection, improved efficiency of renewables, enhanced distribution network operations and  faster energy infrastructure maintenance.

Legacy grids are passé. A large part of the future is bundled up in the huge amount of data which smart grids are generating right now.