Machine Learning Techniques Reduce Uncertainty in Long-Term Performance Reference
EWC Weather Consult, a German pioneer in the optimisation of weather data, has developed a long- term correction method for wind measurements giving far superior results. By using machine learning processes EWC has created a method that successfully minimises yield uncertainties. The new method makes it possible to use non-linear corrections, and by doing so the error in the yield estimates on a wind time-series can be reduced to only 3% on average, even for very complex sites. This is half the error level achieved using the matrix method and one-fifth of the error associated with sector-based linear regression in site assessments.
By Jon Meis, Managing Director, EWC Weather Consult, Germany
EWC Weather Consult, a German pioneer in the optimisation of weather data, has developed a long- term correction method for wind measurements giving far superior results. By using machine learning processes EWC has created a method that successfully minimises yield uncertainties. The new method makes it possible to use non-linear corrections, and by doing so the error in the yield estimates on a wind time-series can be reduced to only 3% on average, even for very complex sites. This is half the error level achieved using the matrix method and one-fifth of the error associated with sector-based linear regression in site assessments.By Jon Meis, Managing Director, EWC Weather Consult, Germany
Training Procedures
NASA’s 3D MERRA data is use as the basis of an individual downscaling to turbine level with a 1 x 1 km² resolution; this information is then combined in the model with the metadata of the location so that it can be corrected over the long term. The atmospheric stability and the typical diurnal cycle of the wind speed, the hub height, orography, vegetation and turbine type all play an important role in the meteorological simulation. For the measure–correlate–predict (MCP) processes the target parameter is the mast measurement of the wind speed, and therefore it is also the point of reference used for determining the systematics of deviations when calculated and observed time periods are compared. Broad and deep neural networks (NNs) are used as underlying machine learning architecture, an approach now commonly used in robotics, character recognition and autonomous driving. The system is trained to perfectly simulate the characteristics of these command variables, or, in other words, the model must be changed so that it approximates to the observed figures as closely as possible. The correction scheme is then applied to the remaining periods of the long-term wind time-series that can be made available, in an ideal world, right up until the previous day.
NASA’s 3D MERRA data is use as the basis of an individual downscaling to turbine level with a 1 x 1 km² resolution; this information is then combined in the model with the metadata of the location so that it can be corrected over the long term. The atmospheric stability and the typical diurnal cycle of the wind speed, the hub height, orography, vegetation and turbine type all play an important role in the meteorological simulation. For the measure–correlate–predict (MCP) processes the target parameter is the mast measurement of the wind speed, and therefore it is also the point of reference used for determining the systematics of deviations when calculated and observed time periods are compared. Broad and deep neural networks (NNs) are used as underlying machine learning architecture, an approach now commonly used in robotics, character recognition and autonomous driving. The system is trained to perfectly simulate the characteristics of these command variables, or, in other words, the model must be changed so that it approximates to the observed figures as closely as possible. The correction scheme is then applied to the remaining periods of the long-term wind time-series that can be made available, in an ideal world, right up until the previous day.
Copied from Nature
Besides high-performance computing architecture the secret of the method lies in a) the algorithms part of the neural networks and b) the input variables. Algorithms attempt to reproduce the structure and process of learning in the human brain and to train the models by rote. As a generalisation, such an ability to transfer insights and patterns to new and different situations is important to people, but it is not usually wanted in networks. They have to be supervised by machines. A random part of the command variable, the met mast data, is always discarded and the successful ability to generalise the results can thus be checked by using independent data. Constant iterative and automatic repetition is necessary until the desired quality of results is achieved.
The event capacity of this training process also depends on the base input, the ‘raw data’. Since meteorological processes take place many kilometres away from the measurement sites, EWC inserts additional information from surrounding areas to the calculation. Thousands of atmospheric circulation factors such as land breeze, mountain valley winds, etc. create a larger input dataset for training. The system then has to select the essential data series efficiently by the so-called feature selection that runs automatically.
Good Prospects for Site Assessments
Today numerical weather models create forecasts, as well as simulations of the three-dimensional atmosphere in the historical context. For this reason the hourly simulation based on the last 35 years of data is constantly adapted to the available measurement data. The results are valuable re-analyses. They provide the wind power industry with hourly wind time-series at hub height as a long-term reference which shows only marginal deviations from the original measurements. This is an appealing idea because establishing the long-term reference at site level has so far been very difficult. By these means the uncertainty of yield forecasts can be reduced by 50% or more compared to the most commonly used methodology, especially in the calculation for the financing-relevant P90. This improves the valuation and financing capability of a wind park project.
The timing of the test and selection of the desired MCP process including free parameters and the associated test can now be eliminated during site assessment. The MERRA data is extended by current weather model analyses so that the long-term data is available right up to the previous day. Remote data can be transferred directly into a long-term reference – without the customer having to wait for the publication of the MERRA data (published only monthly and often delayed by six weeks: too long for the implementation of wind park projects).
Use of machine learning processes also has interesting applications for neural networks making short-term wind power forecasts. For wind power trading the goal is to provide the most accurate power production forecast possible. The benefit for the customer is that balancing costs are kept low and trading margins high – a powerful argument when one considers the sinking management premiums in the trading of wind power in Germany. As the traders use these forecasts for intra-day and day-ahead wind power trading, EWC puts its emphasis on the current day and the next day. These are two forecast regimes that need to be mapped very well by the neural networks, and the current state of the systems plays a crucial role for the next forecast hours. After just a few forecast hours the information is transferred to the forecast from the numeric weather models. This is why real-time data exchange, such as live-production from wind parks, becomes more and more important for top quality forecasts.
Wind is not the only subject for which machine learning processes are relevant. The focus lies on non-linear processes and thus the technique can be used for the optimisation of the atmospheric factors for forecasts used by wind/solar parks and wind/solar park trading portfolios. These special applications are also already established with EWC customers.
About the Company
EWC Weather Consult GmbH is a specialised provider of weather services and forecasts for the energy industry and has been operating since 1999. It is a leading driver of technology for machine learning processes, MCP long-term corrections for yield calculations for wind energy, measurement data processing and customer-specific prediction. EWC provides high-quality weather information, historical weather databases, weather expertise, and short- and medium-term forecasts for wind and solar energy production as well as lightning and climate analyses.
EWC Weather Consult GmbH is a specialised provider of weather services and forecasts for the energy industry and has been operating since 1999. It is a leading driver of technology for machine learning processes, MCP long-term corrections for yield calculations for wind energy, measurement data processing and customer-specific prediction. EWC provides high-quality weather information, historical weather databases, weather expertise, and short- and medium-term forecasts for wind and solar energy production as well as lightning and climate analyses.
Biography of the Author
Jon Meis has been a Meteorologist (German Diploma) and is the Founder and Managing Director of EWC Weather Consult GmbH. He is an expert in boundary layer meteorology, both through his research and from his experience of flying gliders and aircraft.
