OPTIMIZATION OF WIND POWER PRODUCTION FORECAST PERFORMANCE

DURING CRITICAL PERIODS FOR GRID MANAGEMENT

WINDPOWER 2007 Los Angeles, CA

June 3-6, 2007 POSTER PRESENTATION

John W. Zack AWS Truewind, LLC

185 Jordan Rd Troy, NY 12180

jzack@awstruewind.com www.awstruewind.com

1

ABSTRACT

Most wind power production forecast systems are designed to minimize forecast errors over all forecasts and are typically evaluated with metrics that provide a measure of their overall forecast performance. Forecast users often have a higher sensitivity to errors in forecasts made for certain critical time periods or phenomena. There appears to be considerable value in optimizing a forecast system to provide the best overall performance, subject to the constraint that it provides the best possible forecasts during the critical periods. One example of a critical phenomenon for grid operations is a large ramp event in which the wind power production increases or decreases by a large amount over a short time period. The importance of correctly forecasting ramp events will increase as the amount of wind power in the electricity portfolio increases.

AWS Truewind is currently engaged in a research effort to optimize forecast performance for large ramps events on both the hours-ahead and day-ahead time scales. A study of these events in a variety of locations indicates that state-of-the-art forecast systems typically do not perform well for such events because the meteorological processes associated with them have much different characteristics than for the typical cases. AWST is developing a forecast system that uses a phenomenon-switching approach in which the forecast system configuration is automatically switched when the probability of a particular phenomenon is estimated to be high. This paper presents an overview of issues associated with the development and implementation of a forecast system designed for optimal performance in the prediction of large ramp events.

1. INTRODUCTION Users of wind power production forecasts are typically more sensitive to forecast error under particular sets of circumstances. One important example is the sensitivity of grid operators, who must balance load and generation, to large ramp events when wind power production changes by a substantial fraction of the capacity within a few hours. The increase in the amount of wind generation will increase the amount of regulation resources that will be needed on the system to manage such events. The procurement of the right amount of regulation will be key for successfully integrating large amounts of intermittent resources such as wind. The accurate forecasting of large ramp events can facilitate the procurement of the right amount of regulation, which will decrease system costs and ultimately enable larger amounts of wind penetration This paper presents an overview of issues associated with the development and implementation of a forecast system designed for optimal performance in the prediction of large ramp events. Section 2 provides an indication of the frequency of large ramps in power production based on an analysis of one year of power production data from a set wind farms in California. Section 3 presents an overview of the most significant processes that cause large ramp events. Section 4 provides an example of the potential complexity of the processes that cause large ramp events. An overview of a large ramp event forecasting system currently under development by AWS Truewind is presented in Section 5. A summary is presented in Section 6.

2

2. FREQUENCY OF LARGE RAMP EVENTS An understanding of the frequency of large ramp events and its degree of variation between sites is an important part of the forecast system development process for these events. An initial attempt to evaluate the frequency of these events was made with the use of data from a set of wind farms in California.

The percentage of 2-hr periods during 2006 with a change in production equal to or exceeding the value on the horizontal axis for 5 wind farms in the San Gorgonio Pass (SG) area and one each in the Tehachapi, Altamont and Solano areas of CA is depicted in Figure 1. Large ramp events (change of > 50% of capacity in 2 hours or less) at locations in California with have frequencies ranging from about 0.5% to a little over a 4%. The frequency of very large ramps (change of > 75% of capacity in 2 hours or less) varies from slightly under 0.1% to near 0.5% among wind farms examined. The data indicate that, as might be expected, large ramp events at individual wind farms are infrequent and that the frequency of large ramp events varies substantially between regions in California with the Tehachapi and Altamont regions having a highest frequencies and the Solano region having the lowest for 2006.

FIGURE 1. FREQUENCY OF 2-HR RAMP EVENTS BY AMPLITUDE FOR 5 WIND FARMS IN SAN GORGONIO PASS AND ONE EACH IN SOLANO, TEHACHAPI AND ALTAMONT RESOURCE AREAS.

3

3. PROCESSES THAT CAUSE LARGE RAMPS The methods and data required to successfully forecast large ramp events depend upon the atmospheric or engineering processes that produce the event and the look-ahead time. An overview of the main processes that produce large ramp events is presented in the following subsections.

3.1 Passage of Large-scale Weather Systems One source of large ramp events is the passage of relatively intense large-scale weather systems. Large-scale weather systems are the common features that are seen on the TV weathercast or in the newspaper such as low and high pressure areas and cold and warm fronts. The large ramp events are typically associated with the passage of low-level high wind speed features associated with fronts (the boundaries between large-scale air masses) and the associated large-scale storm systems. These wind speed features may only have a modest horizontal scale (hundreds of kilometers) but they typically have relatively long life cycles (several days), are strongly controlled by the evolution of the large-scale atmospheric circulation and predominantly driven by quasi-horizontal atmospheric processes. The combination of long life cycles, fairly large horizontal scale and the strong controlling influence by large-scale weather systems make these features fairly easy to monitor with standard meteorological data and they are typically quite well forecasted by operational physics-based numerical weather prediction models. Thus, the predictability of large ramp evens associated with these atmospheric processes is fairly high in both the hours-ahead and day-ahead modes.

For the hours-ahead forecast time horizon, the key issue in predicting large ramp events associated with these types of weather systems is monitoring and projecting the movement of the large-scale weather systems and the associated low-level wind speed features. The forecasting of development and decay of the features is usually a secondary issue on this time scale. The development and dissipation are more important on longer time scales (e.g. day-ahead). In the hours-ahead time frame, the wind speed features that produce significant ramps can typically be well-monitored with near-surface and other standard meteorological data. NWP models typically provide good forecasts of these systems for up to several days ahead and often also provide very good guidance of the expected evolution of the features on hours-ahead times scales as well.

3.2 Onset of Local or Mesoscale Circulations A second type of atmospheric process that can produce large ramp events are associated with the onset of local or mesoscale circulations. These circulations include features such as sea-land breezes, mountain-valley winds, drainage flows and gap flows. They are smaller in scale than the large-scale weather systems described in the previous section but they are still largely controlled by quasi-horizontal processes. However, their life cycles usually range from a few hours up to a day, which is considerably shorter than the large scale systems.

The much shorter life cycles and smaller spatial scales have twp impacts on the nature of the forecast problem. First, these characteristics make the monitoring of the systems more difficult since their small spatial scales make them hard to diagnose with conventional meteorological data and their short life cycles make it difficult to anticipate the evolution of these systems even when they are diagnosed. For these type of systems, the forecasting of the development or

4

dissipation of the system is often more important that the prediction of the movement of the system. In the hours-ahead mode, some skill in the forecasting of the large ramps associated with these systems can be achieved by monitoring the evolution of these features with surface-based meteorological data. This approach yields better results when the scale of the features is large enough to be adequately resolved by the standard network of surface weather observing sites. A potentially more effective way of monitoring and predicting these phenomena is through the use of remote sensing devices (e.g. Doppler radars) that provide a high resolution three-dimensional view of the wind and temperature structure of these systems. In the day-ahead mode physics-based NWP models are the best tool for forecasting large ramp events associated with these types of atmospheric features. These models can provide fair to good prediction of the winds associated with local or mesoscale circulations for one to two days in advance. 3.3 Vertical Mixing of Momentum (Dry Convection) A third type of atmospheric process that can produce large ramps in power production is the vertical turbulent mixing of momentum. The large ramps in power production associated with this process typically occur in two ways. The first type occurs when there is a layer of high wind speeds above a low speed layer in the height range of the turbine rotors. In this situation the turbulent mixing of wind speed between the high speed layer and the low speed conditions at the rotor-level is inhibited by the stability of the air between the high and low speed wind regimes. However, under some conditions a small decrease in stability or an increase in vertical wind shear can result in a sudden onset of turbulent mixing which brings the high-speed air down to the turbine rotor level. Many of the most explosive upward ramp events are associated with this process. A second type of event can occur when there is layer of high wind speeds from the surface up to a height well-above the top of the turbine rotors accompanied by fairly vigorous vertical turbulent mixing associated with relatively low thermodynamic stability. In this situation a sudden cooling of the near-surface layer can drastically increase the stability of the near-surface atmospheric and decrease the amount of turbulent mixing. In this case, the wind speed can experience a significant and sudden decrease below and in the rotor layer. This can result in a large drop in wind power production over a relatively short period of time.

The salient attributes of this process are that it is dominantly a vertical process that is characterized by a very short and often highly variable life cycle. The key to predicting these events is the anticipation of changes in vertical turbulent mixing associated with changes in wind shear and thermodynamic stability. The magnitude of the turbulent mixing is sometimes very sensitive to small changes in wind shear or stability and hence it is sometimes extremely difficult to predict the changes in turbulent mixing even one or two hours ahead.

As noted previously, the primary tools for hours-ahead forecasting are the monitoring of the relevant atmospheric features and the use of the monitoring data to project the movement and development or dissipation of the features. However, turbulent mixing events are very difficult to monitor with data from meteorological towers since these events are inherently a vertical process. The meteorological tower data typically provides very little information about the vertical structure of the atmosphere above the wind farm’s turbines. Since meteorological towers are the main source of hours-ahead forecast information for most sites, these events are typically not very well forecasted. Significant improvements in the hours-ahead forecasting of this type of large ramp event will probably be associated with the use of remote sensing devices that provide information about the three dimensional structure of the wind sheer and stability below, within

5

and above the turbine rotor layer. Day-ahead and longer forecasts of these types of events are even more difficult. Physics-based NWP models, which are the primary tool for day-ahead and longer forecasts, have little skill in predicting specific occurrences of turbulent mixing events. However, these models generally do provide good predictions of the environmental conditions that are favorable for these events. Thus, they can provide a good indication of time windows when the probability of this type of event is higher or lower than average but not exactly when or where an event will occur.

3.4 Thunderstorms (Moist Convection) Another type of atmospheric process that can produce a large ramp event is a thunderstorm. Thunderstorms are actually a subset of a type of atmospheric phenomena known as moist convection. They are related to dry turbulent mixing processes described in the previous section, but in the case of moist convection, phase changes of water play an important role in the process. The addition of the water phase changes typically results in a process with longer life times, larger spatial scales and more of a horizontal component than the dry turbulent processes. However, the life cycles are still relatively short (an hour to perhaps 12 hours for longer-lived convective systems). Thunderstorm can occur in isolation and may only be large enough to impact a single wind farm or they might be organized into mesoscale convective systems that may be large enough to have an effect on all the wind farms in a region within a short period of time.

The key to successful forecasting of this type of ramp event is the anticipation of the locations of development and dissipation and the wind characteristics of thunderstorms or systems of thunderstorms and the direction and speed of their movement between their development and dissipation, Forecasts of thunderstorm-driven large ramp events are heavily dependent on remote sensing technologies that can track the evolution of thunderstorms. Doppler radar and satellite imagery are currently the best tools for this task. These provide a fairly good indication of the current location, movement and amplitude of these systems. However, the rapid evolution of these systems and the incomplete information provided by the remote sensing systems typically make it difficult to project their evolution even a few hours ahead. Networks of meteorological towers typically provide little useful predictive information for this type of event but they can sometimes be useful in confirming or supplementing the picture provided by the remote sensing tools. The skill in day-ahead thunderstorm event forecasting is mostly tied to the skill of NWP models. As in the case of the dry turbulent mixing events, current NWP models have little skill in forecasting the timing and location of individual thunderstorm development and dissipation but the NWP models do have considerable skill in forecasting the environmental conditions that are favorable for thunderstorm system development and the environmental factors that control the characteristics of the thunderstorms that do develop. Thus, while it is extremely difficult for NWP models to forecast individual ramp events associated with thunderstorms, the NWP models can provide a day-ahead indication of when the probability of large ramp events is above or below average over several hour time windows.

3.5 Reaching Turbine Overspeed (Cut-out) Threshold The preceding discussion of the ways in which large ramps in power production can occur have focused on wind speed changes between the cut-in speed around 4 m/s to the speed at which the turbines reach their generation capacity, typically around 11 or 12 m/s. Large ramp evens that

6

occur in this wind speed range require a substantial change in wind speed and hence a fairly significant atmospheric event. However, large ramp events forced by any of the previously discussed processes can also occur when the wind speed exceeds the turbine cut-out (overspeed) threshold, which is typically around 25 m/s for most modern turbines. In this type of event, very small changes in wind speed can result in a large ramp event if the wind speed is near the overspeed threshold. This type of event is usually very difficult to forecast since small errors in the wind speed forecast can result in enormous errors in the power production forecast.

Obviously, the key to forecasting this type of large ramp event is the prediction of the critical wind speed changes when the wind speed is near the overspeed threshold. The ability to do this varies significantly depending on the type of atmospheric processes that are responsible for the event. Events that are predominantly controlled by large-scale atmospheric processes can often be anticipated with a fairly high degree of reliability one to two days ahead, especially if the NWP models indicate that the wind speeds will go from somewhat below the cut-out threshold to well above the threshold. However, events which are mostly attributable to the onset of local or mesoscale circulations, dry turbulent mixing or thunderstorms are typically very difficult to forecast even a few hours ahead due to the short life cycle of these phenomena and their rapid evolution.

4. COMPLEXITY OF LARGE RAMP EVENTS: AN EXAMPLE As part of AWST’s effort to develop a forecast system customized for large ramp events, AWST has identified and analyzed a number of large ramp events with the objective of understanding how they occur and determining what types of data and forecast models are needed to accurately forecast them. Some of the forecasting issues can be illustrated by examining the large ramp event that occurred in the San Gorgonio Pass during the evening and night of March 22-23, 2005. The regional power production for the period is shown in Figure 2. There was a downward ramp of about 270 MW between the hours of 6 PM and 8 PM PST. This is followed by an upward ramp of 250 MW during the 4-hour period between 9 PM and 1 AM PST, which featured a 1-hour increase of 200 MW between 11 PM and midnight. The available anemometer data from locations throughout the San Gorgonio Pass area as well as data from other meteorological sensor systems were analyzed to understand the processes that led to this ramp event. The data from three anemometers in San Gorgonio Pass is illustrated in Figure 3. Two of the anemometers are located on the eastern side of the Pass where most of the capacity is located and the other anemometer is in the “upwind” central portion of the Pass, which has much lower wind production capacity. Both of the anemometers on the eastern end of the Pass experience a sudden decrease in wind speed beginning at about 6 PM. The speed of the westerly wind drops from over 20 m/s to under 5 m/s in less than 2 hours. During the same period the anemometer in the central portion of the Pass exhibits only a slight decrease in speed that is well within the amplitude of the fluctuations that occur before and after this time. The obvious question is what could have caused the sudden decrease in the wind speed in the eastern portion of the Pass while the wind speeds remained fairly high in the central part of the Pass? Since there is very little data other than that from meteorological towers in the Pass, it is useful to examine the output from physics-based numerical models for this case. The simulated wind speeds at 50 m AGL (upper panel) and at 1500 m above mean sea level (lower panel) for 9 PM PST 22 March from a model run with a 5 km grid cell size initialized at 4 AM PST on the

7

morning of 22 March are shown in Figure 4. These depictions indicate that the model has done a reasonably good job of simulating the large difference in wind between the central part of the Pass (SGC) and the eastern part of the Pass (SGE) that was observed in the anemometer data. The model also indicates that the there are wind speeds in excess of 25 m/s at a height of 1500 m above sea level which is approximately 1000 m above ground level in the central and eastern part of the Pass. Thus the wind shear over the eastern part of the Pass at this time is quite large while the shear over the central part of the Pass is not as large. The output of the physics-based models in combination with the anemometer data in the Pass and the data from the automated surface observing systems (ASOS) at nearby airports and Doppler radars in southern California allow one to construct a reasonable scenario of atmospheric events associated with the large downward and upward ramps in this case. A schematic depiction of the critical atmospheric features associated with this event is shown in Figure 5. The combination of all of this available data suggest that high winds speeds over 20 m/s were present at a few hundred meters above the Pass throughout the event. Prior to 6 PM the stability of the lower atmosphere was near neutral and there was a substantial amount of mixing of the high wind speed air downward to near the surface and the 50 m wind speeds were relatively high in the central and eastern part of the Pass. At approximately 6 PM an area of rain moved into the Pass. The evaporative cooling associated with the rain falling through the relatively dry low level air caused the near surface temperature to decrease. This caused the boundary layer to stabilize and drastically reduced the amount of turbulent mixing. The impact was greater on the eastern end of the Pass because it is at a lower elevation and is further from the source of high speed air several hundred meters above the Pass. The eastern end of the Pass may also have experienced a greater amount of evaporative cooling and stabilization because the low level air was drier prior to the onset of the rain. The result is that the vertical turbulent mixing is almost completely cut off over the eastern portion of the Pass and the wind speed collapses. A substantial amount of turbulent mixing is maintained in the central portion of the Pass and the wind speeds do not decrease much in this area. Since most of the wind power capacity is in the eastern portion of the Pass, the regional power generation drastically declines during this period with only the plants in the central part of the Pass generating much power during the 7 PM to 9 PM period. The rain stops shortly after 9 PM and the near surface air warms slightly as the evaporative cooling associated with the rain ceases. The combination of the boundary layer becoming less stable and an increase in the wind speeds a few hundred meters above the surface due to an increasing pressure gradient causes the vertical turbulent mixing to abruptly resume over the eastern part of the Pass. The wind speeds rapidly increase between 10 PM and Midnight and there is a large upward ramp in the regional power production. The analysis of this event indicates that important physical processes were operating in the vertical direction. Hence, upstream wind data was of no value in anticipating this event as indicated by the fact that the upstream anemometer in the central part of the Pass provided no indication that the winds would suddenly decrease and then abruptly accelerate downstream in the eastern part of the Pass. The output from physics-based models provided some clues about processes associated with this event but errors in timing and the exact magnitude of the wind shear and thermal stability made it difficult to anticipate these ramp events from that data alone. The ability to monitor the boundary layer stability and the wind shear in the lowest kilometer of the atmosphere with remote sensing systems would provide valuable short-term predictive information that can supplement the clues in the physics-based model output.

8

FIGURE 2. AVERAGE HOURLY REGIONAL POWER PRODUCTION (MW) FOR CALIFORNIA’S SAN GORGONIO PASS WIND RESOURCE AREA FOR THE PERIOD 1200 PST 22 MARCH 2005 TO 0600 PST 23 MARCH 2006.

FIGURE 3. WIND SPEED DATA AT A HEIGHT NEAR 50 M AGL FROM THREE DIFFERENT ANEMOMETERS IN SAN GORGONIO PASS FOR THE 1200 PST 22 MARCH 2005 TO 0600 PST 23 MARCH 2005 PERIOD.

9

FIGURE 4. SIMULATED WIND SPEEDS AT 50 M AGL. (TOP PANEL) AND 1500 M ABOVE MEAN SEA LEVEL (BOTTOM PANEL) FOR 9 PM 22 MARCH 2005 FROM A PHYSICS-BASED MODEL RUN INITIALIZED AT 4 AM PST 22 MARCH (17 HOURS EARLIER).

10

Figure 5. A schematic depiction of the atmospheric features that have been diagnosed to play an important role in the ramp event of 22 march 2005. The yellow line indicates the high speed (~20-25 m/s) flow at 500 m to 1000 m above the Pass.

5. A SOLUTION: LARGE RAMP FORECASY SYSTEM AWS Truewind is engaged in a project to improve the forecasting of large ramp events. The direction of the development effort is being driven by a couple of key considerations. First, large ramps are infrequent and unusual events and are generally outliers in data samples used for the training of statistical forecast models for general wind power production forecasts. Therefore, these events typically receive little weight when constructing forecast relationships from a standard training sample. This suggests that a system specifically designed to forecast this type of event is needed to achieve the highest possible forecast performance. Second, a number of different processes cause large ramp events. An effective forecast will likely have to be formulated to estimate the probability of the different type of ramp events.

The new AWST forecast system is based upon a three-step procedure. In the first step, the probability of the occurrence of a ramp rate above a specified threshold amplitude is calculated for each type of ramp event for a target time window. The second step examines the occurrence probability for each type of ramp event that is calculated in the first step. If the occurrence probability for any type of event is above a predetermined threshold, an estimate of the amplitude probability distribution for that event is calculated. In the third and final step the probabilities for each type of event are combined into an estimate of the overall probability distribution for a ramp event above a specified amplitude.

The system is currently under development and testing for California. Preliminary results show considerable skill over ramp forecasts implied by standard wind power production forecast systems but indicate, as expected, that some types of ramp events (e.g. those caused predominantly by vertical atmospheric processes such as the sudden onset of turbulent mixing events) are much more difficult to forecast than other types.

11

6. SUMMARY Large ramps in wind power production are a significant concern for grid operators. Users of wind power production forecasts are typically more sensitive to forecast error under particular sets of circumstances. One important example is the sensitivity of grid operators, who must balance load and generation, to large ramp events when wind power production changes by a substantial fraction of the capacity within a few hours. The future increase in the amount of wind generation will increase the amount of regulation resources that will be needed on the system to manage such events. The procurement of the right amount of regulation will be key for successfully integrating large amounts of intermittent resources such as wind. The accurate forecasting of large ramp events can facilitate the procurement of the right amount of regulation, which will decrease system costs and ultimately enable larger amounts of wind penetration

An analysis of the frequency of large ramps in California indicated that large ramps are a relatively infrequent event at most wind farms with ramps of over 75% of capacity in 2 hrs or less occurring in much less than 0.5% of the 2-hr periods. An examination of the causes of large ramp events in California indicated that large ramp events can be caused by a number of different processes. Optimal forecast performance is likely to be achieved using a multi-scheme forecast system in which individual schemes are formulated for each significant type of process that generates can generate a large ramp event. Finally, a forecast system designed to forecast the probability of ramp events in specific time windows and their likely characteristics (amplitude etc.) is under development for California by AWS Truewind.

7. ACKNOWLEDGEMENTS The contributions of Dan Meade, Joe Nocera, Steve Young, Ken Waight and other scientists at MESO, Inc. and AWS Truewind to the extensive analysis of the large ramp cases, the identification of the meteorological processes associated with the ramp events and the development of innovative approaches to the forecasting of large ramp events is greatly appreciated.