Forecasting Electricity Demand
Read this article. The authors suggest that underestimating demand for electricity services can result in shortages that impact productivity and overall economic growth. Box 1 presents a case study of electricity forecasting in Armenia. How do you think your personal energy consumption helps or hurts your country's economic development?
How should electricity demand be forecast?
Both methodology and data availability make forecasting electricity demand a challenge for experts
There are various forecasting approaches, models, and methods: econometric analysis, general equilibrium economic models, and engineering or mathematical methods – many of them also data intensive. They capture economic and population growth, variations in input prices, and prevailing weather conditions and patterns. Moreover, while the ultimate goal is to forecast electricity demand, in most countries only electricity production data are available for forecasting purposes. As electricity is a nonstorable and poorly tradable commodity, output is a reasonable proxy for total final consumption. But using electricity output data will lead to downward-biased forecasts in the presence of market distortions and correspondingly high suppressed demand.
Ideally, demand data should be broken down across customer categories, such as residential and nonresidential demand, which may behave quite differently. A significant time series of 10–20 years is usually needed, ideally with quarterly resolution to capture seasonal patterns. When good data like these are available, practitioners can employ econometric methods to deliver accurate forecasts (see box 1 on page 6 for a recent example from Armenia).
When data are of lesser quality or do not exist, practitioners may resort to simple heuristics – for example, assuming that electricity demand will continue to grow at a predetermined rate (most frequently, the historical average) or that it is somehow proportional to projected GDP growth.
As an alternative to these ad hoc approaches, we present econometric forecasts based on a "horse race" benchmarking of popular time-series econometric forecasting methods. First, a number of electricity-demand forecasts are computed based on a time series of electricity production in 106 developing countries from 1960 to 2012. The econometric analysis includes the most popular time-series model families. Altogether, 33 model specifications are estimated for stationary electricity production data series; 36 model specifications are estimated for nonstationary series. These econometric specifications comprise both univariate time-series models, where historic electricity production is the sole determinant of future demand, and multivariate time-series models, which include other predictors, such as countries' GDP and demographics. The accuracy of the resulting electricity production forecasts is then evaluated for each country. Specifically, econometric models are tailored to the historic series, leaving out the final five and ten years for model validation. Sample econometric forecasts are then compared against the actual values of electricity output in the validation sample; the best-performing models, which yield the lowest forecast error, are selected.
Time-series econometric forecasts are shown to systematically outperform heuristic methods of demand forecasting; they also perform quite well relative to more refined microeconometric models. Table 2 compares the accuracy of best-performing forecasting models against two benchmark heuristics as follows: (i) electricity demand grows at the same constant rate as in the past, or (ii) electricity demand grows proportionally to GDP growth. The comparison between the time-series econometrics and the heuristics is based on two measures: the "accuracy improvement" or mean percent difference between model average within sample forecast errors, and the "statistical significance," or the percentage of cases for which the difference in model forecast errors between the two approaches is statistically significant. Table 2 demonstrates that forecasts based on time-series econometric methods significantly outperform projections based on either of the ad hoc methods used by practitioners, although the advantage diminishes slightly as the forecasting period is extended from 5 to 10 years. The massive accuracy improvement over the benchmark model that assumes that electricity demand is proportional to GDP may at first seem counterintuitive, given the high popularity of this approach among energy practitioners. But, as shown above, GDP growth is not always strongly correlated with electricity demand. And when it is, multivariate econometric time-series models estimate a more accurate GDP-energy multiplier when compared with a simple 1:1 ratio.
Figure 4. Time-series econometric forecast errors by country categories
"Time-series econometric forecasting methods yield accurate forecast predictions for the majority of developing countries".
Time-series econometric models produce accurate forecasts, although certain types of countries are systematically more difficult to forecast. Figure 4 plots the average electricity-output forecast errors based on the best time-series model across country regions, income categories, electric system capacity, and oil-export orientation. Time-series econometric forecasting methods yield accurate forecast predictions for the majority of developing countries, with a mean average error of 6 percent over the five-year forecast horizon. This implies that, for many countries, time-series econometric methods perform well relative to more data-intensive econometric methods (as illustrated in box 1). The quality of electricity-production forecasts diminishes, however, for the countries of Sub-Saharan Africa, low-income countries, and countries with low electricity-access rates and small electricity-generation systems. The forecasting accuracy of time-series methods is also greatly diminished for countries that have recently undertaken major investments (e.g.,Ethiopia, Cameroon, and Myanmar) or disinvestments (e.g., Lithuania) in electricity-generation assets; countries that have volatile electricity production or rely heavily on electricity imports (e.g., Albania, Benin, and Botswana); and countries affected by war (Iraq, Libya, Syria) or disaster (Haiti). For these countries, the application of more rigorous forecasting methods is advised.
For a country-by-country table showing both historic demand growth trends (2000–15) and projected demand growth rates for 2015–20, see table 3 at the end of this note. The table also provides information on the preferred time-series forecasting model selection for each country, as well as the mean forecast error. Practitioners may find this helpful both as a quick check at the country level on forecasts of electricity demand, or as a first approximation when no other information is available. Those wishing to develop their own econometric models may benefit from indications showing which specifications proved most successful for each country. Graphic representations of each country's historic electricity-demand trend and forecast will also be available in Steinbuks (forthcoming 2017).
Box 1. Forecasting electricity demand in Armenia
Armenia recently underwent an intensive debate about investment in electricity generation – particularly whether an expansion scenario based on an oversized nuclear power plant should be preferred over one based on gas-fired generation. A robust forecast was needed to avoid economically unjustified investments.
Forecasting electricity demand in Armenia is greatly facilitated by high-quality historical data, including quarterly series for aggregate income (GDP) and end-use electricity prices. As the electricity sector is a relatively minor part of the country's GDP and the changes in prices are primarily supply-driven owing to changes in costs (regulated tariffs), one can plausibly assume the exogeneity of explanatory variables. The least squares regression approach was therefore employed to forecast electricity production. Quarterly dummy variables were included to capture the seasonal changes in electricity demand. Separate models were estimated for both residential and nonresidential categories of consumers. The preferred models were selected based on which performed best out-of-sample. Each model was then back-tested using historical quarterly data for 2003–10. The models were evaluated based on the Root Mean Square Error (RMSE) for the forecast years. The model with the lowest RMSE was selected and then refitted for all available quarters (2003 to 2010).
The first table below shows the outcome of the estimated residential model. Overall, the best model explained 91.6 percent of the total variation in residential demand for the period 2003 to 2010. GDP and seasonal dummy variables were found to be statistically significant. The resulting income elasticity is 0.31.
The second table shows the outcome of the estimated nonresidential model. Overall the model explains 91.7 percent of the total variation in nonresidential electricity demand for the period 2003 to 2010. All included variables are found to be statistically significant. The estimated elasticity for income is 0.38; for price, -0.38.
With accurate and reliable electricity demand forecasts in hand, the government of Armenia has gone through several iterations of an investment plan for least-cost electricity generation. It decided to reconsider its earlier approach to investing in excessive generation capacity.