Leveraging High Performance Computation for Statistical Wind Prediction

This paper presents a new application of a particular machine learning technique for improving wind forecasting. The technique, known as kernel regression, is somewhat similar to fuzzy logic in that both make predictions based on the similarity of the current state to historical training states. Unlike fuzzy logic systems, kernel regression relaxes the requirement for explicit event classifications and instead leverages the training set to form an implicit multidimensional joint density and compute a conditional expectation given any available data. The need for faster, highly accurate, and cost-effective predictive techniques for wind power forecasting is becoming imperative as wind energy becomes a larger contributor to the energy mix in places throughout the world. Several approaches that depend on large computing resources are already in use today; however, high performance computing can help us not only solve existing computational problems faster or with larger inputs, but also create and implement new real-time forecasting mechanisms. In wind power forecasting, like in many other scientific domains, it is often important to be able to tune the tradeoff between accuracy and computational efficiency. The work presented here represents the first steps toward building a portable, parallel, auto-tunable forecasting program where the user can select a desired level of accuracy, and the program will respond with the fastest machine-specific parallel algorithm that achieves the accuracy target. Even though tremendous progress has been made in wind forecasting in the recent years, there remains significant work to refine and automate the synthesis of meteorological data for use by wind farm and grid operators, for both planning and operational purposes. This presentation will demonstrate the effectiveness of computationally tunable machine learning techniques for improving wind power prediction, with the goal of finding better ways to deliver accurate forecasts and estimates in a timely fashion.