Forecasting wind speeds at tall tower heights within Missouri

Abstract

Forecasting of wind speeds is necessary for the planning and operations of the wind power generating plants. This research investigates the short term forecasting of wind speeds at tall tower heights for stations within Missouri: Columbia, Neosho and Blanchard. The first objective was to characterize the chaotic nature of this parameter using mono and multi fractal analysis using the Rescale Range Analysis (R/S Analysis) and the Multifractal Detrended Fluctuation Analysis respectively (MF-DFA). It was determined that the system was fractal and there were no trends indicative of increasing fractality and complexity with increasing height. The second objective was the qualitative and quantitative chaotic characterization of the wind speeds using phase-space portraits and the Largest Lyapunov Exponent (LLE) respectively. The methods confirm the results of the fractal analyses. A simple non-linear prediction algorithm, Empirical Dynamical Modeling (EDM) was then used to forecast the wind speeds using a moving window. It was determined that the EDM was comparable to persistence. It beats this benchmark model in the very short term range of one time step or 10 minutes. The third objective was to cluster the data using Self-Organizing Maps (SOMs), having identified the optimum number of clusters as 4 using the Elbow and Silhouette Methods, among others. Three continuous intervals belonging to a particular cluster, which represented approximately 50 percent and over of the input vectors or rows from the data frame were identified. These intervals were then used as inputs into a Long Short-Term Memory Network (LSTM) with variables, pressure and wind speeds, as well as a lagged series LSTM with embedding dimension, d, and time delay (tau). These were compared to the Moving window Auto Regressive Integrated Moving Average (ARIMA) and to persistence. It was determined that the lagged series LSTM improved on the LSTM with wind speed and pressure series inputs, and all models beat persistence. The lagged LSTM beats the Moving ARIMA for at least 2 of the forecasting times of 60 and 120 minutes for all intervals.