Models and Methods of Wind Speed Forecasting

Wind power energy has a significant importance in the electric grid. As it is an inexhaustible and freely available resource, its usage is globally increasing. It is replacing all other costly exhaustible resources that were used so far. So, accurate prediction of wind power is required. It is assumed that by 2020, wind energy resources are going to supply approximately 12% of total world electricity demands [1]. The output of a wind power farm is dependent on various factors, including wind speed, terrain, topographical features and farm layout. And it is highly correlated to the wind speed. So, In order to forecast the wind power, first wind speed needs to be predicted. Wind speed characteristics such as, randomness and uncertainty makes the forecasting of wind power a tedious task. The increasing importance of wind power prediction encourages researchers to develop more techniques to predict wind speed and wind power so that the accuracy can be improved. Wind speed is usually treated as a meteorological time series with uniform distribution.

Forecasting can be done for different time periods [4]:

• Immediate short-term forecasting
• Short-term forecasting
• Long-term forecasting

The immediate short-term is forecasting for few hours ahead, short term is for few days ahead and long-term for multiple days ahead. Here, short-term forecasting is done. The majority of literature regarding time series forecasting can be classified into following types of models: physical model, statistical model, neural network model and hybrid models.

Physical Model

The physical prediction method consists of some physical-based equations to convert meteorological data from a certain time, to the forecasted wind speed at a site considered. This method is considered as an effective way for long-term prediction. The famous physical wind forecast modela are the Pridiktor and Casandra.

Statistical Model

Statistical models are simple and is observed to consistently outperform other models when making very short-term (1 to 3 hours ahead) forecasts (Giebel et al., 2003). These methods take past values and sometimes other explanatory variables into account for making predictions.

Several works are done on statistical models

• Kalman filters have been shown to provide better forecasts than persistence models for forecast periods less than an hour ahead (for example, Kalman (1960), A. Bossanyi, E. (1985)[1]. and Louka et al. (2008)).Kalman filters have also been used with other models (ANN)to improve the accuracy [2](2008)
• Several researchers have shown that ARMA models and the variation in ARMA are useful for time series forecasting (wind speed forecasting). Infact, most of the Statistical model belongs to the AR and MA family.
• ·For example, Brown et al. (1984), Torres et al. (2005), and Nfaoui et al. (1996) accounted for the non-normality and seasonality of wind speed by transforming and standardizing wind speed observations and then fit ARMA models to the transformed data. Several modifications have also been done for example, In [] paper (2016), frequency decomposition is introduced in which the wind speed data is splitted into high frequency and low frequency. Shifting and limiting is introduced in addition differencing which results in better performance.

  ·

• In 2014, ARIMAX model was used in detection of intoxicated drivers.In this paper, data was collected using vechile based sensors, then the data was processed to determine impairment level of the driver.ARIMAX outperformed in comparision of other models.

Bivona et al. (2011) used a seasonal autoregressive integrated moving average (SARIMA) structure to model transformed wind speed values. Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models have been widely recommended due to their simplicities and demonstrated abilities to forecast volatility (Brownlees et al., 2011);

• In 2015, GARCH model is used to predict day ahead electricity price.The GARCH model outperforms a general time series ARIMA model when volatility and price hikes are present.
• Second, Liu et al. (2011) applied the above-mentioned GARCH models for wind speed volatility and found that the volatility of wind speed has the nonlinear and asymmetric time-varying properties.
• Other authors have introduced time-varying coefficients into regression models to forecast wind speed (e.g. Reikard (2008)) or regime-switching models in space and time (e.g. Hering and Genton (2010)).
• A DAR model is proposed which shows that it outperforms even not sufficient historical data is present.

ANN: Artificial neural networks (ANNs) approach has been suggested as an alternative technique to time series forecasting and it gained immense popularity in last few years. The basic objective of ANNs was to construct a model for mimicking the intelligence of human brain into machine [13, 20]. The AI approach is also an effective way to forecast wind speed data (WSD). The advantage of the AI method is to predict future times series data without any predefined mathematical models. In the relevant literature that has been published recently, artificial neural networks (ANNs) have been widely used as a method of prediction and function approximation in nonlinear systems. Many researchers have applied ANNs for time series prediction of climatic variables in different time scales with satisfactory results compared to the traditional techniques.