Trend triplet based data clustering for eliminating nonlinear trend components of wind time series to improve the performance of statistical forecasting models

Time series forecasting techniques suffer from various issues. One of the significant issues that affect the performance of the forecasting model is the variability present in the components of the time series. The conventional forecasting techniques reduce the nonlinearity by eliminating its non-stationary components. However, improvements in results are often not significant. This paper proposes a wind speed forecasting approach capable of handling the variability present in wind time series components. In this approach, first, nonlinearity in wind time series is eliminated using trend-based clustering, and then statistical methods are applied to develop the forecasting models. The classical time-series data clustering techniques build a small number of clusters based on the distance metric for data values. As time-series data exhibits a serial correlation between subsequent observations, the distance metric cannot group similar data without affecting serial correlation. The proposed method presents novel time-series clustering techniques applied on wind data using triplet (Growing, Decline, and Identical) analysis of trend components to identify similar trend shapes. Once the clusters of identical trend shapes are generated, a cluster is selected based on the Pearson Correlation Coefficient. After that, statistical models (ARIMA, SARIMA, GAS) and the neural network-based model (LSTM) are applied to forecast the selected cluster. The proposed models have superior prediction accuracy than the existing forecasting model, as evidenced by the comparison result curve between actual and forecasted wind speed, the histogram of Friedman ANOVA mean ranks, and the performance measuring indices. The proposed models (C-ARIMA, C-GAS, C-SARIMA, and C-LSTM) enhance the performance of ARIMA, SARIMA, GAS, and LSTM models, and the percentage improvement is 39.66%, 24.16%, 21.02%, and 54.38%, respectively.