Do We Need to Implement Multi-Interval Real-Time Markets?

• Author: Biggar, Darryl R.

ABSTRACT

Many market-based power systems have implemented a form of 'look-ahead dispatch' which simultaneously solves for the optimal dispatch and prices over several intervals into the future. A few papers have pointed out that the dispatch outcomes which emerge from look-ahead dispatch may not be time consistent. We emphasise that this time inconsistency is not inherent in look-ahead dispatch but is a consequence of the assumption of linear cost and utility functions, which is arguably a special case. Various augmentations to the dispatch process to resolve the time inconsistency problem have been proposed, but these augmentations suffer from the drawback that they do not allow the power system to efficiently adjust to new information. We query whether it is necessary to implement multi-interval real-time markets. We show how under certain assumptions, a sequence of one-shot dispatch processes will achieve the efficient outcome.

Keywords: Look-ahead dispatch, Pricing, Real-time market, Dual degeneracy

1. INTRODUCTION

   Around the world, as the penetration of intermittent generation and distributed energy resources increases, concerns are being raised about the ability of wholesale electricity spot markets to deliver adequate reliability and efficient dispatch outcomes. (1) In particular, faced with the risk of increasingly large swings in the supply/demand balance there are growing concerns about the ability of 'one-shot' economic dispatch processes to efficiently dispatch the available generation and load resources in the face of ramp rate constraints and other short-term inter-temporal constraints. (2)

   For example, the famous California 'duck curve' (3) highlights how, with increasing penetration of solar in a power system, conventional sources of generation will likely be required to ramp up quickly as solar generation declines in the late afternoon. But some generation resources may be limited in how quickly they can increase their output. If these ramp rate constraints are forecast to bind in the future, it may be efficient to pre-position some resources in the power system ex ante to reduce the cost of adjustment to a new supply/demand balance ex post. (4) In fact, such pre-positioning may be essential if power system reliability is to be maintained. This has led some markets to consider introducing separate procurement of ramping capability or flexible operating reserves. (5)

   But how should such 'pre-positioning' be achieved? Many authors have argued that, in the face of binding ramp rate constraints, the wholesale market should not compute the efficient dispatch outcome for a single period at a time (which we will refer to as 'one-shot' dispatch). Instead, the dispatch process should compute the optimal dispatch over several dispatch intervals at a time, taking into account potential short-term inter-temporal constraints such as ramp rate constraints or short-term energy limits. (6) In fact, as set out in the online appendix, many wholesale power markets have already implemented such practices. These policies are known by various names in the literature including "dynamic economic dispatch" (7), "multi-interval real-time markets" (MIRTM) (8) or "look-ahead dispatch" (LAD) (9). Hua et al. (2019) explain the justification for MIRTM as follows:

   MIRTM allows for more efficient dispatch of generation to meet projected system conditions by pre-positioning resources to cope, for example, with large ramps in net load. In particular, a satisfactory short-term forecast for renewable generation can allow LAD to meet high ramping requirements due to generator variability in a reliable and economic manner. Consequently, the New York Independent System Operator (NYISO) and California ISO (CAISO) have already implemented MIRTM, while the Electricity Reliability Council of Texas (ERCOT) has proposed the approach. In practice, in those jurisdictions which have implemented look-ahead dispatch, the dispatch process forecasts the market state (demand, supply, and network conditions) several dispatch intervals into the future and then computes the optimal dispatch and prices for the current and forecast market states. The dispatch and prices from the first interval in the sequence are used for dispatch and settlement purposes. All the future prices are merely advisory. In this paper we will refer to the first price in the sequence of prices from the dispatch process as the spot price. All the other prices (which, to repeat, are merely advisory) we will refer to as forecast prices.

   However, as conventionally implemented, such look-ahead dispatch processes suffer from two problems:

1. First, as a few authors have pointed out, the prices that emerge in the look-ahead dispatch may not be 'time consistent': Even if the forecasts of the future market state are accurate, the profile of spot and forecast prices which are forecast at the outset may not be the same as the sequence of spot prices that arise as we step forward in time. As a consequence, the generation and load resources may not have an incentive to comply with the dispatch instructions. This is a potentially serious problem as resources may not respond to dispatch instructions (including pre-positioning), potentially undermining system security and reliability.

2. Second, the conventional look-ahead dispatch task only forecasts a single market state in each near-term future dispatch interval. This approach does not lead to efficient dispatch outcomes (including pre-positioning) in the real-world context in which new information can arrive over time about market demand, supply or network conditions in the future. If, for example, there is uncertainty about future wind generation conditions, an efficient dispatch in the face of ramp-rate constraints should take that into account. This is not possible in the conventional formulation of look-ahead dispatch.

   In this paper we explore the significance of these problems and suggest a solution. We conclude that, while the first problem has attracted attention in the literature, it only arises in what is arguably a special case (the case of linear cost and utility functions). The second problem, however, has no practical solution. But, we point out that, under reasonable assumptions, implementing look-ahead dispatch is not necessary to achieve efficient dispatch (including pre-positioning) of the power system. Instead, a single one-shot dispatch process (with no requirement to forecast future market states) can achieve the efficient outcome, provided market participants can effectively forecast the future prices in different scenarios.

   Ramp-rate constraints and their impact on electricity market design have long been discussed in the electricity market literature. For example, Han et al. (2001) model ramp rate constraints in the formulation of what they call 'dynamic economic dispatch'. Several papers have explored the application of look-ahead dispatch in the context of specific wholesale markets. For example, Yu et al. (2005) discuss the implementation of a multi-interval economic dispatch model for the Ontario real-time electricity market; Price and Rothleder (2011) discuss extending the dispatch horizon and look-ahead unit commitment in California's energy market; Xu and Howard (2013) consider a proposal for look ahead security-constrained economic dispatch in ERCOT.

   Several papers assess the theoretical benefits of look-ahead dispatch. Xie and Ilic (2008) discuss the potential benefits of applying model predictive control in the context of a multi-interval real-time economic dispatch problem. Using a model that includes stored energy resources, Zhu and Chen (2012) show that the cost of energy dispatch and the cost of procuring regulation reserve are reduced in their look-ahead multi-interval formulation as compared to the traditional single-interval formulation. Thatte et al. (2014) highlight the cost-saving benefits arising from pre-positioning some generators in response to changes in future dispatch intervals.

   A few papers have highlighted some theoretical issues. For example, Choi and Xie (2016) perturb the Karush-Kuhn-Tucker conditions of the look-ahead economic dispatch to calculate the impact of data corruption on look-ahead economic dispatch results, concluding that the look-ahead dispatch is more vulnerable to data corruption than the static one-shot economic dispatch. Wang and Hobbs (2013) compare the look-ahead real-time economic dispatch model under deterministic and stochastic settings.

   For the purposes of this paper, we are particularly interested in the strand of the literature that argues that look-ahead dispatch or multi-interval real-time markets gives rise to a time inconsistency problem--that the sequence of 'ex post' prices arising from the first dispatch interval in each of sequential runs of the look-ahead dispatch process may not coincide with the sequence of prices forecast by the dispatch process at the outset--even if nothing changes in the power system. These issues are raised by Hogan (2012); Ela and O'Malley (2015); Hua et al. (2019); Guo et al. (2019).

   Such a time inconsistency would be a serious problem. In normal circumstances, wholesale power markets rely on market participants voluntarily complying with their dispatch instructions. (10) If participants expect to face a different sequence of prices in the future than originally forecast, they will not necessarily have an incentive to comply with the dispatch instructions at the outset, undermining the dispatch process, leading to inefficient outcomes and threats to system security. To address this problem Hogan (2016) and Hua et al. (2019) propose extensions or augmentations to the look-ahead dispatch problem. These extensions tie future or subsequent price outcomes to prices that were determined at an earlier time. These mechanisms resolve the time inconsistency problem in the special case of perfect foresight.

   The...