analytical Q Suggest a Link Contact Search Energy


Final published article:
PDF (806kB)
Recommend this page to a friend:
Your name:
Your email address:
Your friend(s):  (separate addresses with commas)
Your message:

Demand forecasting

by Anne Ku (March 2002)

original draft of article published in March/April 2002 issue of Global Energy Business


Post-deregulation, load forecasting is no longer a black art practised by a chosen few. More and more companies are recognizing that getting it right means the difference between staying in business or going bankrupt. Rather than relying on a software package to churn out the forecasts, companies are developing the expertise in-house. For many though, it's still a well-kept secret.


Whether forecasting demand for the next day or next year, it is still very much an art. And like art, no forecast is perfect. In today's volatile power markets, however, imperfection translates to millions of dollars. So it pays to be as accurate as possible.


There's a new saying that goes like this, "a trader is only as good as the load forecast he is provided with." Traders at American Electric Power (AEP, Columbus, OH, get forecasts that are continually refined and updated by the mathematicians and statisticians in the market analysis group. It was one of the earliest parts of the wholesale operations built in recognition for its importance in providing a competitive edge.

A poor forecast could mean the difference between buying power at $15/MWh or $80/MWh; or selling power for nothing in the UK. When the New Electricity Trading Arrangements (NETA - see box) was implemented last year, system prices regularly fluctuated between a low of negative $360/MWh (to sell to the system, in other words, to pay to get rid of your electricity) to an excess of $70/MWh (to buy from the system). Under these conditions, the energy and utilities team at Cap Gemini Ernst & Young calculated that an improvement of 4% in forecast accuracy could save electricity suppliers up to $29 million on balancing costs alone.

Although the accuracy of short term load forecasts (those less than 7 days) has improved greatly over the years, the sensitivity to and the anticipation of unexpected events still defy even the most advanced modelling techniques. Take September 11th for instance. The three minute silence observed in the UK on 14 September at 11 GMT caused the one of the largest drops in system demand- 2,700 MW (7%).

Long term demand vs short term load

According to Nigel B. Lewis, managing consultant at Cap Gemini Ernst & Young's London practice, long term and short term forecasting require totally different sets of skills. Long term forecasting is more about looking at different scenarios such as what my portfolio will look like; what competitors might do; whether next winter will be warmer than usual. In short term forecasting, you have more precise information about your customers, better weather prediction, and you're more concerned about the impact of major events such as television programs which influence demand in 5 to 10 minute intervals.

Under a regulated, monopolistic, public regime, short term (load) forecasting is used to ensure reliability of supply. Long term (demand) forecasting (1 to 5 years) is used to ensure a healthy capacity reserve through long term investment planning. In Brazil, Electrobras is only involved in long range planning studies (10 to 20 years ahead). With electricity privatisation and liberalization under way, load forecasting is now being conducted outside Electrobras.

Prior to NETA, the National Grid was responsible for producing forecasts, as everyone else paid the same price for each half-hour in the UK Pool. Although load forecasting was used in structuring commercial supply contracts, it wasn't until NETA that a financial incentive existed for suppliers to forecast on a regular basis.

For companies which manage full requirement loads, the energy products matched to that obligation are costly. As they are only provided as product offering by selected suppliers, load forecasting becomes a necessary part of deal and price structuring.

Given the high volatilities of electricity prices, it really pays to improving the accuracy of load forecasts. Phil Inje Chang spells out what this means in practical terms (see next article).

Forecast accuracy

"The one thing you can guarantee about forecasts is that they will be wrong," says Lewis. The best that he has seen in the UK is errors of 2%; within any half-hour period. In other words, if the actual load is 100 MW, the best forecast is either 98 or 102 MW, but not closer.

If the inputs to your forecasting model are poor, it would be very difficult to get a good forecast no matter how good your model is. In Brazil, the data recorded on a minute by minute basis are irregular, full of missing points and outliers. The research team led by Professor Reinaldo Castro Souza, of Pontificia Universidade Catolica-Rio, Brazil is developing corrective filters to produce continuous observations as input to the models.

Accuracy of forecasts also depends on type of customer. Lewis observes that residential loads are easier to forecast simply because of the sheer number of residential customers. If one customer does something strange, it has less impact. On the other hand, large industrial customers may exhibit unpredictable usage, such as one that can generate its electricity or operate an extra work shift.

Micro and macro factors also affect the forecast, says Paul Corby, senior VP, Planalytics, Wayne, PA. Commercial usage will be down in a depressed real estate market. If a lot of offices have space to rent then this means there is no one occupying them to use power. In a state of an economic recession, industrial usage will be down because if no one is buying their goods then they will cease to manufacture. The reality is that there is more elasticity in the industrial and commercial sectors than there is in the residential sector.

According to Corby, there is virtually a one to one correlation between weather and residential power consumption. People are constantly on the search for the Holy Grail when it comes to better weather forecasting.

Follow the weather

So it makes sense to get good weather forecasts - otherwise you will have garbage in, garbage out. But most people take weather forecasts for granted. Lewis notes that nearly all electricity suppliers take UK's Meteorological Office's weather forecasts verbatim in producing their own load forecasts. They don't look at the risks associated with weather forecasts such as the uncertainty and subjectivity around the forecasts. Temperature and other numerical data get fed into the models without translating the probabilistic qualifiers which so often accompany these forecasts. "There's a 30% chance of rain, with average temperature around 5 degrees Celsius" gets translated as a single number 5.

One way around this problem is through the use of so-called weather ensemble predictions. Rather than using point forecasts, it makes use of multiple scenarios for the future value of a weather variable. Dr James W. Taylor, Said Business School at Oxford University and Dr Roberto Buizza, principal scientist at the European Centre for Medium-Range Weather Forecasts (ECMWF), Reading, UK found that weather ensembles greatly improve the accuracy of neural network load forecasting. The ensemble conveys the degree of uncertainty in the weather variable.

Living with errors

If no forecast is perfect, what can we do about the errors? Corby advises to adopt a disciplined hedging policy, regardless of which model people choose. One way to hedge the risk of errors is through portfolio diversification, which is explained in Rees and Hooke's article.

Another way is to pass through the errors. Lewis reports one UK supplier trying to incentivize their clients to report their own forecasts as accurately as possible.

More and more, market participants are looking for a better solution to this problem. Instead of spending more time trying to produce better forecasts, they are doing transactions that recognize that forecasts can be far from accurate. Some companies have started to look for energy products that are tied to actual power pool load. Dr Anil K. Suri, chief executive of E-lectrade, Tarrytown, New York, explains that E-lectrade's "percentage of pool" structured energy product was designed to help manage the business risk associated with hard to predict loads.

Final word

Load forecasting shouldn't be seen as a black box or even rocket science. Lewis thinks it's better to start with a simple model, understand it, and improve upon it; rather than start with a complicated one. If a complicated model breaks, you may not be able to fix it. He adds, it's not something you want to outsource because of the need to understand your customers.

And for that reason, it remains one of the best kept secrets - in-house.

BOX: Load forecasting under NETA

The New Electricity Trading Arrangement (NETA) is a bilaterally-traded or OTC brokered market with a balancing mechanism which replaced the UK Pool in March 2001. One difference between the UK Pool and NETA is buyers' submission of load requirements. These buyers include the former Regional Electricity Companies (ex-RECS in their new identities after the wave of mergers and acquisitions), Centrica, small niche suppliers, and internet-based suppliers. Load forecasting is no longer just a concern of the system operator NGC which balances the system but all of NETA's participants.

As Cap Gemini Ernst & Young's Nigel B. Lewis observes, when NETA went live, people were mainly interested in getting something in place. To get the (load) numbers, some people went out and bought software packages. A year on, they've come to realize that they need to know what's inside the black box.

Prior to NETA, the ex-RECS had forecasting capability but no financial incentive to improve. Theirs was more in long range planning - rather than short- which is required to balance the system. The main driver then was to feed into business planning

Post NETA, there are financial incentives to produce the correct load forecast. The penalties of not forecasting correctly translate to a 30% premium on electricity prices. Buyers also face a double-whammy problem: it's most expensive when you need it most. There's a penalty whether you over or under forecast. So some traders take a view, purposely going long or short, but this doesn't avoid the penalty.

Shanti Majithia, Head of Operational Forecasting at National Grid Company (NGC), Coventry, UK observes that the market is usually long as system buy price is more expensive than system sell price. People tend to overforecast. NGC also conducts system load forecast for balancing purposes. If an imbalance occurs, NGC needs to contract for expensive generation.

3 demand charts provided by LCG Consulting.