Chapter 14 - Wind speed forecasting in Nepal using self-organizing map-based online sequential extreme learning machine

Abstract

Wind speed forecasting is crucial for wind farm sustainability, smooth operation, financial feasibility, and selection of suitable wind farm locations. To establish wind farms as a reliable and sustainable source of energy, the nonlinearity nature of wind speed must be addressed. This chapter implements a hybrid Self-Organizing Map-based Online Sequential Extreme Learning Machine (SOM–OSELM) to forecast wind speed at three forecast horizons, that is, short term (6-hourly), daily, and monthly. The study was carried out using historical wind speed for five geographically diverse sites in Nepal. The chapter also implemented benchmark standalone models OSELM, M5, and autoregressive integrated moving average (ARIMA) and another hybrid SOM–M5 model, to compare the reliability of the proposed model. These models were chosen such that OSELM is the absolute nonlinear model, ARIMA is a true linear algorithm, and M5 represents a pairwise linear model. This chapter also highlights the advantage of using a proper data splitting method which is often considered trivial during the model development process. SOM was implemented as a data splitting method which reduced the variance between the test and training datasets compared to the traditional data split methods. The hybrid models SOM–OSELM and SOM–M5 were built on the SOM data partition, whereas standalone models OSELM, M5, and ARIMA were built on the traditional partition. The model’s performance was evaluated using correlation coefficient (r), root-mean-square error (RMSE), mean absolute error (MAE), relative RMSE, relative MAE, Willmott’s index (WI), Nash–Sutcliffe coefficient (NSE), and Legates and McCabe index (E). The study showed that the proposed SOM–OSELM model was found to be more efficient than the benchmark models for all five sites with respect to short-term and daily time periods, achieving a higher correlation (r) and model accuracy (WI, NSE, and E). However, for the monthly forecast horizon, SOM–M5 outperformed the SOM–OSELM model marginally in terms of all performance metrics except for WI. The monthly wind speed was more linear compared to the daily and 6-hourly wind speeds, hence, SOM–M5 being a pairwise linear model outperformed the SOM–OSELM model while forecasting monthly wind speed. The study also concluded that the performance of the proposed model, along with the benchmark models, was independent of the geographical locations of the study sites. From the study, it was observed that SOM–OSELM and SOM–M5 models performed better than their standalone models. Thus it was concluded that the SOM data partition method is an efficient data splitting method for wind speed forecasting. Although SOM–M5 outperformed the SOM–OSELM model to forecast monthly wind speed, the hybrid SOM–OSELM model was observed as a robust and reliable tool to forecast wind speed at 6-hourly, daily, and monthly horizons. Therefore this chapter recommends that the SOM–OSELM can be utilized as a decision support system by the wind farms to study the nature of variation in wind speed at 6-hourly, daily, and monthly intervals. This will assist the wind farms for real-time grid operations, scheduled maintenance, maximize the power reserve for time periods with lower wind speeds, and eventually establish themselves as a reliable and sustainable source of energy.