EFFECT OF TIME HORIZON ON WIND SPEED PREDICTION WITH ANN

Proper utilization of renewable energy sources in electricity production is inevitable due to the environmental concerns and global warming fight. Therefore, predictability of renewable electricity is a very significant issue for a long time. Main aim of this study, different from the literature, is to investigate the change of wind speed prediction errors for different time horizons. Different prediction time horizons (10, 30, 60, 90 and 120 minutes) were used, and the results were compared through the error measures and the regression values. The mean squared errors and the regression values vary between 0.819 and 5.570, and between 77.8% and 97.1%, respectively. The prediction error changes almost logarithmically, and the rate of change decreases with the increasing time horizon. A new analysis approach was proposed to see the change of the prediction error with time horizon. The equation, y = 1.5413ln(x) 2.7428, representing the change of the mean squared error with time horizon was obtained.

proposed a new approach to predict hourly wind speed characteristics [30]. Akinci and Nogay proposed a wind speed estimation method which uses data of neighboring measurement station [31].
Time horizon is a significant issue in wind speed predictions as studied by others. In study of Foley et al., effect of length of the time horizon on wind speed and wind power predictions was presented [42]. It was mentioned that wind speed prediction error increases with the increasing length of the time horizon. Wang et al. studied on ANN approach to predict wind speed [43]. They achieved to get better results in predicting wind speed at different time horizons. It is noted that larger prediction periods reduce the accuracy.
Advances in applied mathematics offer some other methods like MARS, CMARS, RCMARS, GPLM, etc. for real life problems on prediction [44][45][46]. In the near future, these methods will make it easier and accurate.
About wind speed prediction, most of the studies in the literature aim to predict wind speed better than the previous attempts. Main aim of this study is to evaluate the prediction performance of the ANN approach at different time horizons but not to develop the best ANN model to predict wind speed. Therefore, different from the studies in the literature, the present study focuses on the change of the prediction error with time horizon. The evaluated time horizons in the analyses are 10, 30, 60, 90 and 120 minutes. The results were compared through the error measures and the regression values. Results of this study may be used to understand accuracy and sensitivity of the wind speed predictions.

PRESENTATION OF THE DATA
The data used in this study is from a measurement station located in the Aegean region, Turkey. At 30 m measurement height, average wind speed, maximum wind speed, minimum wind speed and standard deviation were stored in ten minutes time interval. The data from 2002 to 2005 were used in the analyses.
Monthly mean wind speed changes were presented in Figure 1 for each year. The highest wind speed values occur in summer months. Maximum monthly mean wind speed is 9.21 m/s (July 2004). As can be seen from the figure, monthly mean wind speed values change significantly. This high variability makes wind speed prediction difficult.

ARTIFICIAL NEURAL NETWORK APPROACH
Today, the ANNs are being used in many different studies and applications. The ANN based methods are usually used to forecast the short term data. In the present study, same configuration of the ANN was used to evaluate the performance of the ANN on short term wind speed predictions at different time horizons. Figure 2 shows the ANN structure used in this study. As input parameters average wind speed, maximum wind speed, minimum wind speed and standard deviation were used in the present study. The wind data used in the analyses were 10 minutes measured data. Additionally, values of year, month, day, hour and minute were used as input parameters in the ANN model. Therefore, the number of input parameters is nine. The neural network used in this study is a two-layer feed-forward network with sigmoid hidden neurons and linear output neurons. As a training algorithm, Levenberg-Marquardt back-propagation algorithm is used. Number of hidden neurons is selected as 40.

Figure 2. The ANN structure
The data of the first three years were used for training, validation and testing issues in the ANN method. 70% of the data was used in the training stage, 15% of the data was used in the validation stage, and 15% of the data was used in the testing stage. The data were divided up randomly for the stages.
Additionally, the random data from the last year were used in the additional tests to see the efficiency of the ANN. While evaluating the results of the ANNs, two different approaches were employed as the mean squared error (MSE) and the regression value (R value).

RESULTS AND DISCUSSION
The results of the ANN method were given in this section. Firstly, the initial results were given for the training, validation and testing stages. Secondly, the results of the additional tests were presented.
The initial results of the MSE and the regression analyses were presented in Table 1 for each time horizon. The MSE values vary between 0.819 and 5.570, and the regression values vary between 77.8% and 97.1%. The increasing length of the time horizon increases the MSE and decreases the R value. The increase of the MSE decreases with the increasing prediction period. In Figure 3, scattering of the predicted and the actual wind speed values were presented. The scatter can be understood by the R value. The lower R value means more the scatter. Therefore, the scatter becomes more as the prediction horizon increases. The highest overall R value was calculated for 10 minutes time horizon as 0.970. The additional test results were given in Table 2. The additional tests were performed by using the data from the last year. 106 data were taken randomly, and the prediction tests were done for the same time horizons (10, 30, 60, 90 and 120 minutes). The MSE values for 60, 90 and 120 minutes time horizons decreased comparing to the initial results.   In this part, some comparisons were done with other studies which have similar conditions. Sfetsos compared the performance of three forecasting models for two different locations to predict wind speed values of one hour ahead by using past hourly data [17]. The achieved root mean squared errors (RMSEs) for the ANN model were 1.200 m/s and 0.724 m/s (one hour ahead prediction) which are lower than the results presented at the additional tests (RMSE=1.860 m/s for one hour ahead) of this study.
Abdel-Aal et al. proposed using of abductive networks to forecast mean hourly wind speed at Dhahran, Saudi Arabia [22]. The R values achieved in the mentioned study were 0.826, 0.853 and 0.842 for different models. The R value (0.861) attained in this study (one hour ahead prediction) is better. The change of the MSEs at different time horizons was presented in Figure 9. The errors change almost logarithmically, and the rate of change decreases with the increasing time horizon. The equation fitted for the MSE eases to calculate the error for any time horizon. For instance, the MSE for 1440 min (24 hour) ahead forecast is 8.466. The representation of the change of the MSE with time horizon may be an important approach for the forecasting studies. However, in the literature, there is not much study using this approach.

NN
Neural network  Artificial neural network