Classification of tall tower meteorological variables and forecasting wind speeds in Columbia, Missouri

Abstract

The wind speeds given in 10 min intervals is forecast using multiple methods inclusive of persistence, statistical methods of ARIMA as well as artificial intelligence methods of Artificial Neural Networks. Tall tower meteorological variables in Columbia, Missouri are clustered using Self-Organizing Maps after the optimal number of clusters was determined using the Elbow and Silhouette methods among others. The optimal number of clusters, k was given as 4 for all methods. The data were then grouped into three Intervals which consisted of approximately 50 percent and over of vectors or rows from the data frame. These intervals were then used as training and testing for the forecast models of Long Short-Term Memory Networks with pressure and wind speeds as inputs as well as lagged wind speeds as inputs. Other models using these intervals in our analyses include Moving Autoregressive Integrated Moving Average (ARIMA) and persistence. From the results obtained from the ARIMA, the metric of the root mean square error (RMSE) ranged from approximately 0.6 to 1.0 ms−1 for forecast horizon 2 to 12 in increments of 2. Interval2 had the upper and lower values and thus showed most variability in errors because it encompassed most of spring, all of summer and the beginning of fall. The moving ARIMA showed lower errors than the LSTM with pressure and wind speeds inputs for all the intervals. This may be attributed to the difficulty in representing the system’s non-linearity and high dimensionality by using just the wind speeds and pressure as inputs. The lagged co-ordinates of the wind speed was then examined and used as inputs for the LSTM. The metric used for the evaluation of prediction of the forecast horizons of 60, 120, 180, 240, 300 and 360 min or 1, 2, 3, 4, 5 and 6 h ahead is the Normalized Root Mean Square Error (NRMSE). These models were compared to the benchmark model of persistence. It was determined that all of the models beat persistence and the LSTM with the lag series outperforms the LSTM with pressure and wind speed as inputs. The Moving ARIMA is now beaten by the lagged series LSTM in all intervals for at least 2 time forecast horizons of 60 and 120 min or 1 and 2 h. It is thus shown that the Artificial Neural Network method with the lagged series inputs is the best performing model.