Modeling Multi-horizon Electricity Demand Forecasts in Australia: A Term Structure Approach.

• Author: Hum, Stan

1. INTRODUCTION

   The National Electricity Market (NEM) in Australia was established in December 1998 and comprises five regions: New South Wales (NSW), Victoria, Queensland (QLD), South Australia and Tasmania. The NEM is both the physical power system that operates in and connects these five regions as well as a wholesale electricity market. The NEM is one of the longest power systems in the world and reached a generating capacity of 65,252 megawatts in December 2021 and traded AUD 11.5 billion in the financial year 2020-2021. Wholesale trading in the NEM is conducted as a spot market where supply and demand are instantaneously matched through a centrally-coordinated dispatch process which is managed by the Australian Electricity Market Operator (AEMO). Consequently, AEMO is responsible for ensuring sufficient electricity to be generated, traded and dispatched to meet consumption (AEMO, 2021). Wholesalers bid capacity and retailers buy electricity from the wholesale grid at the current spot price and sell to consumers at a heavily regulated price. As part of its operations AEMO generates one-day ahead electricity demand forecasts for the NEM that are updated every half hour until the time of dispatch. These forecasts play an important role in informing the decision-making process of many electricity market participants. Consequently, this paper is primarily concerned with the evaluation and interpretation of these official forecasts of electricity market demand.

   Forecasts of electricity demand are used by market operators for scheduling and dispatch of generation capacity which is crucial to system stability (Adams et al., 1991; Harvey and Koopman, 1993; Cottet and Smith, 2003; Li et al., 2017; Zhang et al., 2018). For generators, demand forecasts are an important driver of strategic choices involved in bidding and rebidding of capacity, whereas for retailers demand forecasting affects decisions about the balance between hedging and spot acquisition of electricity. One of the more influential papers on electricity demand forecasting in the Australian context is Hyndman and Fan (2010) who provide a comprehensive semi-parametric approach to the forecasting problem. Clements et al. (2016) provide a more conventional time series approach. (1)

   An empirical characteristic of official AEMO forecasts is that they appear to over-predict demand consistently for longer forecast horizons while gradually reducing the degree of over-prediction as the target time interval approaches. This property is highlighted in Figure 1 which plots the smoothed median forecast for each half-hour horizon for the NSW and QLD regions of the NEM. The smoothed estimate given by the dashed line is based on locally estimated scatter plot smoothing (or lowess smoothing) developed by Cleveland (1979). The forecasts are arranged from the longest forecast horizon of 16 hours (32 half hours) to the shortest horizon. Although this phenomenon has been observed before in energy markets (Auffhammer, 2007), in oil and output markets (Nordhaus, 1987), as well as in macroeconomic forecasting of real GDP and inflation (Batchelor, 2007; Ager et al., 2009), no particular attention has been paid to interpreting it. (2)

   To understand the features of the NEM forecasts in Figure 1 a model is developed that explicitly takes account of the information provided by forecasts generated at all time horizons. A special property of the model is the identification of a set of latent factors describing the evolution of demand forecasts over time. The factors can also capture how forecasts are updated as a result of the arrival of news which changes shorter-term forecasts relative to longer-term forecasts. Using the Australian data, the empirical results provide evidence of a three-factor term structure model. The three factors are interpreted as level, slope and curvature factors, a result that is akin to the factor structure of models of the term structure of interest rates. See, for example, Diebold and Li (2006). The dominant factor is the level factor, with contributions from the slope and curvature factors. The level factor captures the persistence in multi-horizon forecasts with similar weights attached to the forecasts at each forecast horizon. The slope factor captures the updating of forecasts as the time horizon approaches the dispatch time, while the curvature factor reveals that information common to very short-term and very long-term forecasts are important in improving forecasts, as is information common to intermediate-term forecasts.

   In addition to providing a valuable interpretation of the official demand forecasts, a further advantage of the term-structure approach is that it allows for a more efficient use of the official forecasts to be developed. The adoption of a factor structure generalizes the concept of pooling forecasts proposed by Bates and Granger (1969) and Granger and Ramanathan (1984). It is well known that pooling several forecasts has several optimality properties, as well as having been found to generate superior forecasts in practice (Timmermann, 2006; Elliott and Timmermann, 2016). However, pooling typically involves the adoption of a single weighted average of forecasts rather than using multiple factors which is a feature of the term-structure multi-horizon forecasting approach proposed here. Note, however, that it is not the objective of this paper to provide a complete forecasting model for electricity demand, but rather to provide a novel interpretation of the existing official forecasts and a way of using them more effectively.

   The rest of the paper proceeds as follows. Section 2 provides the necessary institutional details of the NEM forecasting framework followed by a discussion of the statistical properties of the multi-horizon electricity demand forecasts. A term-structure model is specified and estimated in Section 3 resulting in three factors being identified. In Section 4 the economic value of AEMOs official forecasts is evaluated in terms of a cost-loss decision model (Richardson, 2000; Foley and Loveday, 2020). Concluding comments are given in Section 5.

2. STATISTICAL PROPERTIES OF NEM FORECASTS

   The NEM trading day is divided into 48 half-hourly intervals with the first interval starting at 04:01 and ending at 04:30, and with the 48-th trading interval starting on the next calendar day at 03:31 and ending at 04:00. This setup is illustrated in Figure 2 which shows how trading days r and [tau] +1 straddle the same calendar day. From an operational perspective, generators must place their initial bids for trading intervals during the day of [tau] + 1 before 12:30, which is known as the gate-closure time. Any bid after this gate-closure time is considered as a rebid. Rebids allow generators to adjust their initial bids in response to new information, such as changes in consumer demand. AEMO provides demand forecasts to generators every half hour starting from the gate-closure time until the targeted trading interval. This means the targeted trading intervals during a day have different numbers of demand forecasts available. For the first half-hour trading interval between 04:01 and 04:30, there are 32 available forecasts. For the last trading half-hour interval between 03:31 and 04:00, there are an additional 48 half-hourly forecasts less one for a total of 32 + 48 - 1 = 79 forecasts made available sequentially. In all the empirical analysis of Sections 3 and 4 the number of forecast horizons is set at 32 as this is the minimum number of forecasts that are available for all target trading intervals.

   The forecast data used here consists of official forecasts of half-hourly average demand and realized historical demand for NSW and QLD. The number of trading days is 1491 from Saturday 2 July 2011 to Friday 31 July 2015, resulting in a total sample size of 1491...