St Lucia Tesla Battery Q & A

A lithium solution was chosen for a few reasons, including its energy density (which enabled more storage within limited space), and speed of response.

We are utilising flow batteries for other applications such as our Heron Island Research Station microgrid project

Our experience found that this is an emerging area, where standards etc. are still catching up and it’s difficult to get clear guidance on regulatory requirements in a commercial setting.

It was one reason why the ultimate site for the project was chosen to be outdoors and adjacent to a solid concrete block wall with no habitable space behind it  (as opposed to windows etc.).

The Powerpack had a few key advantages over other options at the time of our procurement (which notably occurred around mid-2018 – we acknowledge the market has evolved since then).

Foremost it had great energy density in a modular package that was aesthetically pleasing (as opposed to ‘shipping container’ style solutions).

This was important considering our desired location and its constraints for layout, as well as it being visually prominent on campus.

The fully integrated and ‘turn-key’ nature of the Powerpack system was also an attractive aspect, with learnings from previous projects highlighting the importance of this.

Using the battery as part of a ‘living laboratory is a key goal of the project alongside its commercial drivers.

It is located prominently in the engineering precinct of campus, with a large live dashboard display adjacent to engage passers-by about the battery.

This dashboard is also available online

Academics have also begun to incorporate various aspects of the battery into their teaching programs, and it has already attracted interest from interstate and international delegations.

Some of the battery’s revenue streams are unique to a behind-the-meter asset (like peak demand lopping), as well as to UQ’s specific circumstances being a spot price exposed customer.

As discussed in the report, different customers will see different value from this (such as sites with higher demand charges).

As a behind-the-meter asset though, there are services that the UQ battery does not currently participate in that would potentially represent value to other projects, such as the Regulation FCAS market.

The challenges and learnings from UQ’s battery are largely applicable to any battery project in the NEM at a high level.

The key point is that any revenue stream which can be leveraged through forecasting and predictive control can be easily accommodated in the DRE control framework.

The battery is controlled by UQ’s in-house developed ‘Demand Response Engine’ or DRE.

This uses AEMO actual and forecast price data to maximise the ‘spread’ between charge and discharge prices subject to the constraints of the battery system.

There is not a fixed charge or discharge setpoint, but rather DRE relies on the cyclic nature of the energy market and places all control in the algorithm to continuously look forward to optimise how the battery should behave over the forecast horizon to maximise revenue streams.

The pros and cons of this approach, and ideas for helping to improve it, are discussed further in the report.

The minimum spread the battery requires before it sees profitable arbitrage potential currently is selected based on:

1. The marginal usage cost of the battery (i.e. what fraction of the capital cost does charging/discharging the battery utilise with respect to the systems nameplate lifetime throughput) and,
2. The minimum ‘operating cost’ caused by the ancillary charges you allude to.

This minimum spread threshold has typically been set in the $40-$50/MWh range.

This was based on NEM price forecasts from several consultants that UQ commissioned for use in a range of projects (including the Warwick Solar Farm) in order to forecast the average daily spread we could expect, as well as applying factors to account for assumptions such as battery uptime etc.

Any arbitrage optimisation problem is highly non-linear, furthermore, the physical and operational specifications of the battery define several linear and nonlinear constraints.

This frames a rather complex optimisation problem to solve and find a true global minimum however the relatively short and discretised forecast horizon (~24hrs x 30min resolution) works favourably in generating an accurate result with relatively low latencies in compute time.

Due to these factors a Sequential Least Squares Quadratic Programming algorithm is currently sufficient with respect to both time complexity and solver performance.

The optimization problem is framed within a receding horizon control framework and as such, near real-time data is fed as input every time new financial information is received (approx. every 5 minutes).

This is possible due to the fast compute time relative to the sampling interval. As the DRE portfolio expands, the optimisation problem will become more complex at which point this approach will be re-assessed.

This is a central area of focus for future improvement and optimisation.

At this stage we don’t have any fixed ideas as to how to improve the forecasts we rely on and will be experimenting with a range of techniques and indicators.

Machine learning will play some part, likely for factors outside of price (e.g. demand, generator availability, weather etc.)

We anticipate we will be undertaking significant back-testing to help try and inform effectiveness of different approaches, whilst also continuing to monitor actual performance and tweaking our approach in real time as we go. Watch this space!

This is not an analysis that we had time to perform prior to finalising the Q1 report, as it involves delving quite deeply into each five-minute forecast snapshot and planned control reaction in the lead up to each cap interval.

It’s something we hope to look closer at in the future. Based on intuition and anecdotal observation alone, we would say that missed coverage is a roughly even mix between not enough storage left and purposeful exposure due to the current forecast observations and revenue opportunity.

This will vary significantly by asset. For example a chiller may need a five minute step down period from the time of receiving a switch off signal to being off, whilst a diesel generator can take anywhere from 10-15 minutes to grid sync and ramp up to full power output.

This will mean that we need to rely more heavily on forecasting the behaviour of these assets in advance, as they lose a lot of the advantages a battery has of being able to ‘react’ to an unexpected price outcome.

This becomes even more important once we move to 5-minute settlement and highlights the importance of improving our price forecasting abilities.

The theoretical study of this predictive control has not yet been formally published. It is our intention for a whitepaper to be released by the end of the year.

It is probably not accurate to call the control framework in its current state as stochastic in nature.

Whilst there are uncertainties in forecast inputs, the current inability to model this uncertainty makes it unwise to frame the problem with a stochastic control approach.

Instead, we simply assume that the pre-dispatch reports we currently receive are the expected values and thus the problem becomes a plain model predictive control problem.

As mentioned in the report, the ability to enrich the raw predispatch models will hopefully result in a more accurate expected value for the forecast market inputs however this will still culminate in the use of a standard MPC framework at this stage.

As a behind-the-meter asset that is less than 5 megawatts, the battery isn’t required to register as a generator or load with AEMO, and thus does not need to bid into the spot market and follow dispatch instructions.

We are instead able to react to dispatched prices and the AEMO price forecasts as we choose.

As UQ installs more storage however - or the thresholds for registration change - this is something we will likely need to consider in the near future. 

The optimisation is currently single objective in its current form. This is largely because peak demand charge co-optimization has not yet been integrated in production. When this does occur, the current intention is to stage this as a cascaded MPC framework rather than a single multi-objective optimization framework.

The reason for this is largely because peak demand charge optimisation acts on a longer time horizon than arbitrage optimisation (1 month compared to ~1 day). Furthermore, as our high resolution (30-minute) peak demand forecasting currently has a 7-day horizon, the peak demand charge optimisation problem is not fully observable for most of each month.

As a behind-the-meter asset the battery operates as part of the site’s existing connection agreement and is thus subject to the same demand charge regime as the site. T

his varies per site and per NSP, but in our case comprises a mix of volumetric charges (c/kWh), monthly peak charges ($/kVA), and monthly capacity charges set on a rachet basis ($/kVA).

These need to be factored into the calculation of the ancillary operating charges that are discussed in the arbitrage section of the report.  

The battery itself does not have its own MLF or DLF as it is a behind-the-meter asset, but the main NMI for the site does. This needs to be factored into the calculation of the ancillary operating charges that are discussed in the arbitrage section of the report.

As a behind-the-meter asset, the battery did not require a standalone connection agreement but rather an amendment to the site’s existing connection agreement.

We worked with Energex (our DNSP for the site) to achieve this, which was a relatively novel process for them too considering the infancy of commercial scale batteries at the time.

Ergon had clear guidelines outlining the level of modelling required, which in our case was very limited due to the size of the battery and the very strong connection point.

As no modelling was required and size was under registration threshold, AEMO was not directly involved.

This varies greatly though, so specific circumstances should be discussed with each NSP who can provide further guidance.

As seen in the report, passive uses like contingency FCAS are certainly more commercially lucrative than more active uses like arbitrage.

The battery’s business case relies on the stacking of multiple value streams however. UQ is comfortable operating the battery within the limits of its warranted throughput, and assessment of degradation rates will be a key learning that is closely watched.

It’s worth noting that in our planning for implementation of the functionality, we expect peak demand lopping to have a high demand on the battery’s storage capacity (eg. significant discharging required to be effective).

The warranty is based on overall MWh throughput (full charge/discharge cycles), and not the frequency of charging and/or discharging. UQ’s control of the battery manages to these throughput limits as a result – effectively an overall average of one cycle per day. So far our usage has been substantially below this level for some of the reasons discussed in the report.

The Powerpack system enables 100% depth of discharge (to nameplate capacity) so we work on this basis. This is ultimately controlled by the battery management system which provides an overarching set of hard rules which will constrain what DRE can direct the battery to do (temperature de-rating is another example, although not experienced to date).

The warranty is based on overall MWh throughput (full charge/discharge cycles), and not the frequency of charging and/or discharging. UQ’s control of the battery manages to these throughput limits as a result. 

There are several packaged solutions on the market however for the same capital cost and significantly reduced ongoing costs, the decision to build in house was an easy one.

The in-house solution yielded numerous benefits over enterprise packages, the three most significant being:

• Full control over the design and enhancement of control methodologies which are custom to a behind-the-meter spot exposed energy customer,
• Ease of access to data for transparent distribution with students, staff and the wider community and,
• Ability to expand to additional energy assets through in-house resourcing and expertise.

UQ is lucky to have the expertise in-house to develop this system and pilot this technology.

We fully acknowledge that this may be relatively unique to us, and is also something that we can justify with a total energy spend of over $20m per annum.

We hope that by sharing our learnings it will help and inspire others to follow, even if not jumping in as fully as we have – you could perhaps extract 80% of the same value for less effort/resourcing.

We also hope our learnings will help the overall sector grow, develop, and build more capability.

It largely depends on what the ultimate goal is. Functionality like contingency FCAS can be implemented very simply by partners like Enel X.

Basic arbitrage controllers that are more heavily rules based could also be easily coded. The effort comes from wanting to improve and co-optimise these functions.

Our investment in DRE has also perhaps been heavier than would have been the case if it was for a battery alone, as it is the foundation of an overall demand response portfolio.

It’s not an area that we have put a lot of thought to at this stage, as it was primarily developed to help solve an operational need UQ had.

That said, we appreciate our work and learnings could have a lot of value for others, and may look to offer this on an open-source basis once further development has been undertaken.

The DRE control framework was built from the ground up using in-house resources and expertise.

The specification of an external RTU as the LLI for the battery system was chosen around the same time that the prototype Demand Response Engine was being developed.

As such its inclusion was for risk mitigation as much as functionality should DRE not have come into production. The functionality of the RTU encompasses:

• Connection agreement conformance (i.e. nil export),
• FCAS Control
• Lighting Control
• Rate limiting

While the Tesla SMC is relatively flexible with its internal controllable modes it is still not able to replicate all the functionality of the RTU.

This is partly due to the much lower pricing typically seen in the contingency lower markets and the potential impacts of being called to charge at the wrong time (either peak demand, or high spot price) – noting though that DRE could co-optimise to help mitigate these risks.

It is also a function of our partnership with Enel X, who have traditionally been geared up to offer contingency raise only due to the nature of the other assets in their portfolio (who can’t load add/charge).

This is something we keep a watching brief on the merits of though, and may be implemented in the near future. 

At present, we are the only FCAS battery in QLD with data publicly available, so comparisons with others is difficult.

Whilst FCAS pricing across the NEM is often identical cross the mainland regions, Q1 saw some relatively unique circumstances with separation of regions which created highly disparate outcomes region to region.

As such we see limited value in comparing our FCAS revenue to batteries in other regions.

The battery quarantines 10 minutes at full discharge (0.185 MWh) at all times for FCAS purposes.

In theory this does reduce potential arbitrage revenue, but we often find that the battery doesn’t complete a full cycle for arbitrage purposes each day due to other constraints such as a lack of sufficient spread for a sufficiently long duration.

This did impact arbitrage revenue outcomes during volatile intervals in Q1 as the battery sometimes ‘ran out’ early, but even in that quarter’s highest priced interval ($2,295/MWh), this would have only cost $424 at worst case – which is insignificant compared to total FCAS revenue of $44,000 for the quarter. We see this as a worthwhile trade off.

Optimising the balance between storage capacity and rated power is a key commercial consideration for any battery project.

It is definitely correct that a battery with much higher rated power and much lower storage duration would be able to capture more of the currently lucrative FCAS value stream.

There are other constraints that need to be considered though, such as the standard configurations available from suppliers (especially for turnkey solutions like the Powerpack) as well as network constraints (in UQ’s case we were largely limited by available transformer capacity) which may limit the practicalities of designing a system purely to maximise FCAS revenue.

Exploring the overall value stack offered by batteries was also a key driver for UQ, and hence a nominal 1MW/2MWh gave a lot of options for doing this.

The battery quarantines 10 minutes at full discharge (0.185 MWh) at all times for FCAS purposes.

At the moment the inverse is true – if the battery is fully discharging for arbitrage purposes, the battery is opted out of the contingency raise FCAS market (as there is no headroom left to give in the event of a frequency deviation).

This means that UQ may miss high FCAS prices in order to chase spot prices. It is often the case that these two markets mirror each other closely, and when considering UQ’s net share of FCAS revenue (after Enel X fees), it remains more lucrative to seek the arbitrage revenue in those intervals.

In some circumstances this is not correct however, and even with ‘high’ spot prices (eg. $200/MWh) it may be more commercially lucrative for the battery to charge, thus doubling the FCAS capacity bid into the market at an even higher FCAS price (eg. $2,000/MWh).

This is an area we are hoping to add to DRE as a co-optimisation function during the second half of the year.

No – based on current rules as a behind-the-meter asset the battery does not have to pay any charges that a standard load NMI doesn’t have to pay.

As an emerging field, it is not yet clear what the technical parameters required will be to participate in a potential future synthetic/virtual inertia market. It’s an area we are watching closely with interest.  

As a market for this service does not yet exist – as we are a behind-the-meter asset – it’s not something we’ve currently explored with Tesla, but considering our hardware is the same as Hornsdale, there shouldn’t theoretically be any reason this wouldn’t be possible through future controls upgrades.

This is not a functionality we currently plan to implement on this specific battery installation, primarily due to their being no established market to commercialise this service.

It is however something that has been experimented with at our Gatton research battery by the UQ ITEE research group.

Grid forming with batteries is also an area that is being worked on through the Heron Island Research Station standalone microgrid project.