Short-Term Wind Forecasting: Can We Do Better?

Increasing Amount of Variable/ “Intermittent” Power Generation

Final Avatar 80x80-Logo-SG-1-and-2-and-IX-LOGO-e1363114874895-150x150

 

Dominic Geraghty

 

Let’s first discuss how much renewables’ production capacity we can expect.

Wind and solar power generation facilities are being deployed at a steady rate across the nation. Their cost of fuel is zero, and the source of their power is virtually perpetual. And power production from wind and solar resources creates no pollution.

Most states have instituted Renewable Portfolio Standards (RPS) – targets for the percentage of power generation (kWhs) that must come from renewable sources by certain future dates.  The RPS target percentages range from 10% to 40% (Hawaii) and the target dates from 2015 through 2025.

cropped-DSC_0288_2.jpgWind and solar differ from traditional power generation plants in that their output is less certain and not controllable --  there is (1) a diurnal variation – wind is stronger at night on average, and solar is available only during daylight hours, and (2) a short-term variation (intermittency) due to rapidly changing wind strength and cloud cover. In addition, wind strength can exhibit steep, unexpected “ramps” -- both up-ramps (increases in strength) and down-ramps -- which can persist for hours. These ramps occur between 50 and 1,000 times per year.

For this dialog, I am going to focus on wind intermittency.

We’ll leave solar intermittency for another day – then we can discuss some interesting new developments related to very short-term intermittency forecasting based on local cloud-movement monitoring.

Increased “Firming” Of Capacity Needed As More Renewables Are Deployed

To maintain the same target level of regional service reliability when renewables are part of the generation portfolio, power system operators have to increase the level of “reserves”, i.e., the back-up power that is kept available to cover production and demand variations from seconds to minutes to hours.

Since renewable energy production can vary substantially within very short periods of time -- seconds and minutes, an increased amount of fast-reacting reserves is required relative to traditional levels. In competitive wholesale markets, these reserves are bid in and priced on an hourly basis. They are quite “pricey” and include a “regulation” product -- for maintaining the system frequency within a very tight, and unforgiving, sub-minute target range -- and spinning reserve products which can be made available within minutes.

The increased reserves requirement raises the overall cost of power, and this is reflected in the market as increases in the “nodal prices” of regional power grids.

Short-Term Wind Forecasting Models

So, in competitive wholesale power markets, nodal prices are affected by variations in wind production.

Wind farm operators, system operators, and energy traders use a variety of wind forecasting algorithms driven by localized or semi-localized wind data sets. Many purchase wind forecasting services on a regular basis from specialist companies, e.g., 3Tier, AWS TruePower, Windlogics, Weatherflow, Global Weather Corp, Garrad-Hassan.

Some of these wind forecasting service providers focus on the quality of their wind data sets, others emphasize the quality of their algorithms (often an “ensemble” of algorithms from which a “master” forecast is derived). Good data and advanced algorithms are both required for the highest quality forecasts.

Wind data is expensive to collect, especially as the granularity, frequency, and three-dimensionality is increased. There is obviously a trade-off between the increasing cost of more granular data and the resulting accuracy improvement garnered.

In general, the further you go out in time, the more uncertain the wind forecast is. However, the phenomenon of wind ramps represents a highly uncertain near-term “extreme” event with major ramifications for near-term nodal pricing.

To do better in forecasting short-term wind strength/direction, and wind ramps, a “rapid-refresh” model is needed. The quality of the initial conditions that launch the iteration is very important. The model will perform best when continuously initialized with data communicated from local wind sensors, where available, and iterative feedback from its forecasts of the more recent time periods. These models need to be “tuned” to local conditions.

What Is The Accuracy Of Current Wind Forecasting Models?

There has been some very good work done on measuring the accuracy of wind forecasts using “back-casting”/”benchmarking” approaches. There has also been good work done on estimating the value of better forecasts. See the archives for two interesting analyses:  (1) Keith Parks: “Wind Energy Forecasting: What Is It Worth?”, and Ahlstrom, Mark, et al., “Atmospheric Pressure - Weather, Wind Forecasting, and Energy Market Operations”.

A Utility Example: Increased Accuracy Created Substantial Savings

In 2011, the utility holding company Xcel operated about 4.1 GW of wind production in its three utility subsidiaries: PSCO, NSP, and SPS.

DSC_0139_2_2-150x150Xcel developed a sophisticated forecasting system over time which uses six public forecasting models including the NCAR system. As a result, Xcel gradually reduced its average monthly forecasting error* over three years as follows: 2008: 19.4%; 2009: 18.0%; 2010: 14.3%. For NSP, for example, the error was reduced from 15.65% in 2009 to 12.2% in 2010, resulting in about $2.5 million in annual savings. For SPS, the savings were $400K/year, and for PSCO, $3.1 million/year.

The utility estimated that the normalized value of the benefits of a 1% improvement in wind forecasting accuracy was about $50K to $75K per year per 100MW of wind capacity.

The total cost of developing Xcel’s advanced forecasting system was about $900K. Based on a simple payback period calculation, this investment provided a very attractive return to Xcel.

The Xcel results are particularly interesting, and credible, given Xcel’s experience as the largest operator of wind power generation in the U.S. for seven consecutive years through 2011.

Other Analyses of the Benefits of Improved Forecasting Accuracy

An NREL analysis (Lew, Debra, et al., “The Value of Wind Forecasting”) reported a mean average error (MAE)* in the range of 12% - 16% for WECC. Extrapolating the WECC results to the U.S. as a whole, they estimated that a 20% improvement in day-ahead MAE is worth about $260 million/year for 14% wind penetration, and $975 million per year for a 25% penetration rate.

The Ahlstrom et al. paper cited above reported on the value, or realized cost savings, associated with improved wind forecasting as follows:

(1)    For the WECC system with 30% wind and 5% solar penetration, a 10% improvement in day-ahead MAE is worth about $100,000 per year – doesn’t seem like a very large absolute return for such a large system, does it?

(2)    With improved wind forecasting, AESO reported reducing the firming requirement to plus or minus 50MW from the 240 MW firming need predicted by a “persistence” forecast,  for  500 MW of total wind capacity

To put the above numbers in context, the uncertainty related to annual demand/load forecasts today is about plus or minus 3%, compared to wind at about plus or minus 20%. Shorter-term wind forecasts (except for ramping prediction) are better: AESO’s day-ahead wind forecast MAE is 1.3%.

DSC_0316_2 150x150Other benefits associated with better forecasts include the prediction of wind farm curtailment mandates due to transmission congestion and/or off-peak “must-run” conventional generation capacity. Curtailment can create significant losses for wind farms. For example, curtailment in MISO has sometimes been as high as about 33% of available wind capacity (private communication to author). To maybe state the obvious, benchmarking of historical forecasts needs to allow for actual curtailment mandates that have occurred.

Forecasting the curtailment costs requires the integration of the improved wind forecasts with a regional power system/market model -- a topic for a future dialog, depending on the level of the smart grid community’s interest.

* We need to be careful of the measures of accuracy we use (there is the potential for “apples and oranges” comparisons because reported measurements can averaged over different time periods)

So, What Is The Business Opportunity For Advanced Wind Forecasting?

First, Better Forecasting Is Valuable

  1. Day-ahead improvements have been shown to have substantial value even when the forecast is far from “perfect”
  2. The highest potential benefits involve shorter time horizons – up to 12 hours using rapid-refresh models, vertical wind profiles, and more frequent measurements (“rapid cycling”)
  3. Seemingly small improvements can create significant savings, e.g., improved data provided by GE sensors and analytical tools used by First Wind Holdings (WSJ 3/14/2013, page B5) resulted in a 2% increase in power production, which boosted revenue by more than $2 million/year (First Wind’s CEO stated: “the wind business is about nickels and dimes”)

How Big Is The Business Opportunity?

Let’s look at the market size and the expected market growth rate.

Think of the market in terms of two separate segments:

  1. Bringing everyone up to the accuracy of today’s state-of-the-art forecasting: in the U.S., the benefits of this are estimated at about $975 million (WECC study extrapolated), and if assume costs are 33% of the benefits (based on Xcel’s benefit to cost ratio above), then this market segment size is about $325 million
  2. Improving upon today’s state of the art. This will likely involve diminishing returns, i.e., the benefit to cost ratio will decrease at the margin as the rate of growth of costs to achieve the higher accuracy increases relative to the growth of the benefits of the achieved incremental accuracy

It’s difficult to come up with the logic for the potential market growth rate – does anybody have any ideas on how to do this?

Currently, the vast majority of wind forecasting services purchased monthly are priced modestly. No company’s technology seems to have a significant "edge” for the type of forecast data being delivered to date. However, some internal forecasting systems appear to have superior performance, e.g., Xcel’s, cited above.

There appear to be two potential pricing points: a commodity market for fairly run-of-the-mill wind forecasts (low-price competition), and a specialist market (value-priced competition) for customized, higher-accuracy forecasts.

How much should a wind forecasting business invest in developing models/software and on data collection? Weather towers, LIDAR, SODAR, and SCADA systems are all very expensive.

Specialized Opportunity – An Energy Trading “Edge”?

Can a trader develop an “edge”, using proprietary wind forecasting models? What would an “edge” be worth?

An energy trading company can create an arbitrage opportunity if it has private access to a wind forecast that is superior to that which others are using.

In addition to high quality data and advanced algorithms, wholesale market protocols will play a role in optimizing trades. It may be possible to use more sophisticated bidding algorithms, including using option value pricing based on a probabilistic risk analysis (private communication, Keith Parks, GWC).

However, it is not easy for the trader to quantify the potential size of the incremental trading profits in order to come up with a rational basis for the appropriate size of an investment in advanced wind forecasting system and bidding algorithms.  And over time, market protocols will evolve to either decrease the level of imperfect information in the market, or to offer products that reduce the size of arbitrage possibilities.

Therefore, for a vendor in the energy trading market segment that is offering specialized forecasts with integrated power system/market bidding simulations, the benefits side of the value proposition is difficult to quantify, and consequently the selling process has its challenges.

Coming Soon: Some New Information on the Value of Better Forecasts

There is a public/private project underway involving the DOE, NOAA, industry forecast vendors, system operators, and university and government research groups to determine the incremental value of both a new high-resolution rapid-refresh weather forecasting model and additional observations.  Results were originally expected in 2012, but the project report has been delayed. Hopefully it will be published soon.

Conclusion

What we know for sure is that the challenges of accommodating the variability and intermittency will only increase as more renewables are deployed nation-wide to comply with the mandated RPS percentages.

Leave a Reply

Your email address will not be published. Required fields are marked *