An optimisation model will maximise revenues from any price differentials, large or small. Therefore, accurate price forecasts are essential, especially in the presence of small price differentials otherwise, the arbitrage trade may result in a loss. Many studies have used perfect foresight assumptions initially to evaluate their model (Ekman & Jensen 2010; Kanakasabapathy & Shanti Swarup 2010; Drury et al. 2011; Locatelli et al. 2015). While this approach is justified when exploring revenue potential, other studies have taken a more realistic approach by including operational strategies to capture storage value. This section looks at the different storage operating and bidding strategies previous studies have proposed.
Sioshansi et al., (2009) used the past 2 weeks’ prices to derive the storage operation schedule but calculate the value from actual prices, a technique they refer to as backcasting. In their study, the two- week backcasting method yields between 85-90% of maximum possible revenues. This is possible due to price patterns occurring cyclically. At two weeks, weekday-weekend effects are captured and the horizon is sufficiently small to capture seasonal effects but not long enough to capture short term effects such as disturbances which persist on a daily-weekly scale. In a later study, Sioshansi et al., (2011) have used a 1-week backcasting lag; there are no clear criteria for the choice of a backcasting lag. There is uncertainty surrounding the optimal backcasting lag and by shedding light on this problem, more information would be gained on how persistence at different timescales affect storage revenues. Other more advanced strategies have been proposed to capture value; He et al., (2011) argued against co-optimisation and propose storage power and energy capacities are allocated in sequential auctions based on the timescales to actual power delivery. To illustrate this point, they choose a week ahead auction representing a generation company which may not wish to use its peaking plants and hence enters a contract with the storage owner to provide power. This auction is followed by a day ahead auction which represented a trader looking for price arbitrage. Finally, an hour ahead auction is carried out, with a System Operator (SO) who wishes to utilise the remaining capacities for frequency
response. This strategy does not require co-optimisation across markets but manages to allocate storage capacities to the three stakeholders.
This strategy is particularly interesting as the sequential allocation of capacity by delivery timescales is feasible. The authors note that this strategy benefits from neutralising actions whereby commitments from a week ahead auction and those from the day-ahead auction cancel each other. For example, if the week ahead auction requires the storage system to discharge 10 MW at a specific time t while the day ahead auction requires that the storage charges by 10 MW for the same period, these actions cancel each other out and there is an efficiency gain since no charging or discharging occurred as a result.
There is an important aspect which the authors do not address and this remains unclear; based on the sequence of auctions, the bulk of the allocations should take place on the week ahead auction. Implicitly less capacity is available to the day ahead auction and consequently even less so to the hour ahead auction. As the short-term markets generally tend to be more volatile than longer-term ones, there is a possibility that this strategy performs the worst. In the example presented the authors chose a generation company that underutilizes storage capacity at the week ahead auction, shown in figure 2.6.
Figure 2.6: Storage capacity allocation to a generation company at the week ahead auction. Source
He et al., (2011)
With a larger number of generation companies participating at the week ahead auction, less capacity will be available for the later auctions. This raises the question of how to optimally allocate/reserve capacity between the three revenue mechanisms. In the absence of perfect foresight, this is difficult to achieve. In fact, the authors recognise this and state:
“However, the exact amount of storage’s value in each auction depends essentially on the specific case setup and the input data. The numerical results should therefore be interpreted with caution.” He et al., (2011, p.1584) .
Nevertheless, the operating strategy has further potential; if the contracts were non-binding (or have low penalties), the storage owner could utilise each auction mechanism to select the highest revenue service and even reverse previous commitments. Such a strategy could potentially be close to the co- optimised values but in reality, such low penalty conditions and non-binding contracts are not common and hence not explored further in this thesis.
Alternative ways to deal with imperfect foresight have been used; in order to assess the business case for CAES under a profit maximising objective in 2030, Lund & Salgi (2009) are required to forecast the future prices, at an hourly resolution. They use an average expected price of 54 Euro per MWh for 2030, a figure which Danish authorities expect (cited by the Lund & Salgi 2009). The authors then scale this average price to mirror the price variations in 2005, deemed as a typical year. In essence, the hourly 2030 price is an extrapolation of 2005 data.
This approach however fundamentally assumes that the price distribution of 2005 will be similar to 2030 at an hourly resolution. However, with renewable energy generation, demand side response and other changes in demand such as electrification of heat or electric vehicles, this technique cannot be applied within the GB context. Furthermore, inference from average prices are not very useful for storage value; mathematically an average price of 54 Euro/MWh price could simply be a constant price for the whole period or an infinite number of variations of peaks and troughs such that their averages are 54 Euro/MWh. It is specifically the magnitude and frequency of these peaks and troughs that are of relevance to storage arbitrage.
Long term price forecasting has also been used by Yucekaya (2013). The author conducts 100 simulations of hourly prices 30 years into the future using Monte Carlo simulations. The author utilises two random variables, one for the electricity price and the other for the price of natural gas. These are fed into the costs (and revenues) of the model since discharging and charging variable costs depend on the price of gas and electricity respectively. This technique is an alternative to the provision of a short-term operating strategy; its disadvantage, however, is that over such long timescales simulated prices can be substantially different. This challenge is endemic to forecasting methods in general and hence why specific scenarios are used to predict the impact of the changes. National Grid (2015c) for example have modelled energy futures using scenarios of high renewables penetration, electric vehicles…etc.
If a future generation mix (and demand level) is known, it is possible to use the merit order of generation dispatch to simulate market clearing price. Foley & Díaz Lobera (2013) used this technique
to simulate the future prices in 2020, scaling wind penetration. This technique is particularly effective and relevant to markets where prices are determined through a gross pool, such as the Irish market Foley & Díaz Lobera (2013) investigated. To evaluate future prices and storage value, Grünewald et al., (2011) also make use of the merit order of generation dispatch. Storage operation depends on whether the system is long, in which case the energy storage system absorbs excess (thus acting as demand) or short, whereby power is discharged instead meet the deficit and reduce generation from peaking plants. There are clear advantages of using the merit order of generation dispatch to calculate prices – they are potentially more resilient to future changes as they are based on sound principles. By comparison, price forecasting techniques which rely solely on historical data are less likely to fare well under changing market conditions.
Other studies (Walawalkar et al. 2007; Nyamdash et al. 2010) used fixed dispatch strategies for storage operation; the charging and discharging times are chosen so as to match the periods of lowest and highest prices respectively. Fixed dispatch techniques are particularly simple to use and represent a basic strategy for capturing arbitrage revenues.
This section showed that the majority of storage revenue problems have been formulated as LP or MILP problems. In the absence of perfect foresight, however, several approaches were undertaken to estimate prices. These techniques broadly included price forecasting, backcasting, merit order dispatch and fixed dispatch strategies.