How will short-term forecasting work in the U.S. offshore wind sector?
What short-term forecasting lessons can be applied to the U.S. offshore wind market from mature markets?
In early 2022, the U.S. Government released the rights to several offshore regions for auction off the Northeastern U.S. coast. This led to billions of dollars (USD) of investment from key developers and established plans for the first major offshore wind farms in North America. The current U.S. administration is advocating for 30 GW of offshore wind installations by 2030. While it is unclear if the short-term generation forecasting requirements will be the same as their onshore counterparts in these markets, key questions arise on whether the complexity of offshore operations will make forecasting for these assets different or more difficult. I will explain what some of the differences will be, how to apply current practices and address new advances in forecasting technology.
Whereas seasonal outlooks and long-term energy assessments usually focus on a wind farm’s expected performance against typical or average conditions, short-term forecasting is highly impacted by current and near-future (0-72 hour) meteorology. Short-term wind generation forecasting is crucial to many end-users for wind farms of all scales. Owner/operators may use generation forecasts for planning maintenance and scheduling downtime. Grid operators (ISO/TSO) may use them for planning portfolio-wide impact on the electrical grid for real time pricing and reliability. Energy traders who participate in real-time and day ahead markets may use them for advanced warnings of high and low wind events, and how they will impact tradeable information.
In the U.S., short-term wind generation forecasting is now common practice for onshore wind farms, Improvements in meteorological physics, computational power and algorithm sophistication have grown alongside the increasing trend in installed onshore capacity, but the focus on offshore wind farms has so far been limited.
DNV’s short-term forecasting tool, Forecaster, has been predicting wind events for offshore wind farms in the North Sea for over 10 years. Key short-term forecasting lessons have been learned during this time which can be applied to the U.S. offshore wind market.
There are three topics to consider when forecasting for offshore wind farms:
1. Impact of the ocean surface
Onshore wind farms are built amongst uneven or severe topography, with high surface roughness that can impact turbulence and alter wind directions. Even wind farms in flat terrain have frictional interactions that lead to vertical shear profiles of speed and direction from surface to turbine hub height. This effect is quite different for turbines erected over water, which has a much lower friction coefficient than over land.
Numerical weather prediction (NWP) models serve as the basis for most forecasting tools. Most include surface roughness with boundary layer physics schemes. For example, the U.S.-based Global Forecast System (GFS) and the European Center for Medium-range Forecasts (ECMWF) both specify surface cover characteristics uniquely for both land and ocean, depending on the geocoordinates of the wind farm’s grid cell. Therefore, the models inherently calculate different wind profiles at the different types of farms.
Sea surface temperature (SST) can impact the temperature and humidity profiles enough to create a different wind profile over ocean than over land. Careful initialization of SST into the forecast models will create more realistic behavior in the models. Most global models now update SST with satellite-derived data assimilation, which could be increased at higher frequencies with in-house downscaled mesoscale models.
2. Size of the turbines and farms
Turbine size and rating, and the number and spacing of turbines across a wind farm, are very important factors to producing a generation forecast. The so-called power model is the algorithm which converts forecasts of wind speed, wind direction and density at each turbine into an aggregate farm generation forecast.
Onshore wind turbines have grown from approximately 60m to 100m in recent years, with a typical hub height of 80m. However, offshore turbines can be twice as large, and plans for new farms expect much larger installations. Between 2000-2010, many wind farms were on the 100-500 MW installed capacity range, with 50-200 turbines rated in the 2-3 MW range. However, now several offshore farms in operation have 200+ turbines with 5-10 MW ratings, equating to capacities well above 1 GW.
The experience acquired in forecasting across the spectrum of turbine make, model, rating, and count has revealed that the power model is generally scalable to new technologies and larger ratings. At an individual turbine level, the physics are well understood that introducing higher ratings and bigger blades is not expected to impact prediction accuracy.
However, the expanse of offshore wind farms does require extra attention. Several existing farms are so wide that they span multiple NWP grid cells and have micro meteorology subzones where some turbines upwind or downwind of the prevailing wind may be out of phase with other turbines and/or impacted by wake effects. DNV has pioneered and championed the so-called blockage effect on offshore wind farms at such scales, and its forecasting tool has identified ways to subset groups of turbines at these massive sites. Forecasting on smaller groups or even individual turbines has shown a reduction in errors over the One Big Farm (OBF) method.
3. Use of Machine Learning
While traditional incorporation of NWP wind speed forecasts in an OBF approach are conceptually straightforward for onshore and offshore alike, it is well documented that such forecasts lack the necessary accuracy at short-term horizons. Machine learning (ML) optimized by incorporating on-site data can reduce error metrics (e.g., mean absolute error or root mean square error) by several percent. Steady improvement to underlying NWP will continually add small incremental gains on accuracy, but the addition of ML can create a step gain if applied correctly.
ML can use ‘feature engineering’ to add complexity to simplistic forecast algorithms that once depended only on wind speed, direction, and density. Adding more variables, such as SST, ocean salinity, humidity and other seemingly unrelated variables can add higher order impacts that ML can utilize
Finally, training on these large data sets, both historically and in real-time, will become more important for ML forecasts. Access to observational data allows ML to mimic on-site conditions that simplistic power model cannot predict.
Understanding ocean interactions
Offshore wind farms will be a major player in the U.S. energy transition, and while the exact use case of short-term forecasting is unclear, current, and future advances in forecasting products can be adapted for offshore fleets to fit in with onshore markets. Offshore wind farms will pose new challenges to how forecasting is currently performed on onshore farms. The ocean and its interactions with the atmosphere above must be understood, and the increased size in offshore turbines and farms must be considered. The use of on-site data for ML may be useful in quantifying short-term wake impacts and other variables which may impact the wind flow and subsequent generation produced by the site.
While the U.S. offshore sector is still several years away from operating assets, DNV’s Forecaster is well positioned with extensive offshore wind forecasting experience from the European market to assist with the offshore revolution.
Learn more about Forecaster
DNV’s Forecaster delivers insight into renewable energy supply and demand through reliable data-driven predictions to help you maximize profits, avoid missed opportunities and optimize operations.