Short term prediction of extreme winds using artificial neural network and multiple lineer regression

In recent years, the climate change has emerged with global warming and which has become to subject lots of investigation like today. This situation also triggers the number of occurrence of extreme events in meteorological variables such as temperature, precipitation and wind. So new searchings have being conducted in the field of study wind power for pretend damages of extreme winds. This thesis has been prepared at the extend of TUBITAK project with the title of Investigation of Extreme Winds and Turbulence Conditions in the New European Wind Atlas and Development of Estimation Systems'' numbered 215M387. For this purpose, maximum wind data have been obtained from Marmara, Aegean, Southeast and Mediterranean Regions by using eighteen wind farms (WF) of Turkey. The locations of these stations are not clearly stated for privacy reasons. Each of the RT data is processed hourly, because the temporal resolution of the data is different. In addition to mentioned thigs due to WF measurement scale height are differ from each other. Elevations have been transfared same amounts as it possible. The power law method was not used when carrying wind speeds for requested level. This level was determined as 80 m. Therefore, the wind speed of the stations without 80 m measurement data was transferred by establishing a curvilinear relationship between different measurement levels. However, since the measurement levels are different, 80 m data could not be obtained for each station. For example, in a station with measurements of 30 m and 50 m, the wind speeds were first raised to 70 m by means of the curvilinear relationship between these two levels. Then, considering therelationship between 50 m and 70 m wind speeds , 90 m wind speed data was obtained. There are lots of ways to predict wind speed and power in literary. These consist of numerical weather prediction models and statistical approaches. Examples of statistical approaches are Multilinear Regression (MLR), Autoregressive Model (AR), Moving Average (MA), Autoregressive Moving Average (ARMA), Integrated Autoregressive Moving Average (ARIMA), Artificial Neural Networks (ANN), Fuzzy Logic (FL), Genetic Algorithms (GA) and Adaptive Network Based Fuzzy Inference System (ANFIS). In this thesis being used methods are Artificial Neural Networks (ANN) and Multilinear Regression (MLR). ANN is one of the techniques of artificial intelligences. It has been developed by imitating the stimulation and information received by the sense organs through the neurons in the computer environment. Today it is one of the technics of being used almost all computational science fields. ANN can also be thought as a black box which processes given inputs and generates outputs. This system process data in parallel and it also learn correlation coefficient among neurons with the principle of minimalize and renew mistake. In other word ANN uses the method of trial and error. In the real world, a dependent variable is expressed by several independent variables. The method used to explain the relationship between two or more independent variables affecting a variable with a linear model and to determine the effects of these independent variables is called multiple linear regression analysis. The formula of multiple linear regression: Where, Y is the dependent variable that is tried to be estimated, X1, X2, ..., Xm are independent variables, 𝑎 is the regression constant, 𝑏1, 𝑏2, 𝑏3, ..., 𝑏𝑚 are the regression coefficients and finally ε is the error term, a is the regression constant and which indicates the value of the dependent variable when all the arguments are zero. 𝑏 values are also can be call as partial regression or partial slope coefficients. As an input in both methods; Maximum wind speeds at time (t-2), (t-1) and t have been taken and maximum wind speed at time (t+1) has been tried to be modeled. Because the ANN method is sensitive to the architecture that makes it, a sensitivity analysis has been carried out to determine the most appropriate ANN architecture. Therefore different learning algorithms like Levenberg-Marquardt backpropagation, Conjugate gradient backpropagation with Powell-Beale restarts and Conjugate gradient backpropagation with Fletcher-Reeves updates, 2 different activation functions, 4 different hidden layers and 4 different neurons in each hidden layer have been used. Finally, 3X2X4X4 = 96 different ANN architectures are created for each WF and the best ANN architectures are determined for each WF. Then, the results of the analyzes of the wind speeds over 20 m/s, which are determined as both the maximum wind speeds and the extreme wind speeds, were compared with the results of the MLRs. The aim of this thesis is to develop a prediction model based on Artificial Neural Networks (ANN) and Multiple Linear Regression (MLR) by using maximum wind data from eighteen different Wind Farms (WF) located in the Marmara, Aegean, Southeast and Mediterranean regions of Turkey. With the help of these models to be developed, the loss of energy and damage caused by extreme wind is expected to be reduced. The results and suggestions obtained in the thesis can be listed as follows: 1) When ANN architectures of maximum wind speeds are examined; The architects of 1_2_3_4 (trainlm, tansig, 3, 9), 1_1_3_4 (trainlm, logsig, 3,9), 1_1_4_4 (trainlm, logsig, 4,9) and 1_2_2_4 (trainlm, tansig, 2, 9) were obtained as 6, 5, 6 and 1 respectively. As it can be understood from this, all architectures have the trainlm learning algorithm and 9 number of neurons in common. 2) As in the case with maximum winds, there is not a single best ANN architecture for all WF's in extreme winds. The 3_1_1_2 (traincgb, logsig, 1, 4), 1_1_1_2 (trainlm, logsig, 1, 4), 1_2_1_2 (trainlm, tansig, 1, 4) and 2_1_1_2 (traincgf, logsig, 1, 4) has high performance and 4, 3, 2, 2 of these architects were found, respectively. Their common features are 1 secret layer and 4 neurons. 3) In the prediction of both winds (maximum, extreme), when the analyzing the results of ANN and MLR, it was seen that the error of the train was smaller than the test as expected. 4) In terms of ease of use, MLR is more practical than ANN, because MLR is an equation. However, ANN is very complicated in terms of its structure. The advantage of ANN is that it has the possibility of developing ANN through different architectures. Generally, the ANN and MLR results obtained from here are close to each other. 5) In addition, suggesting for later studies is that testing different ANN architectures, using longer datasets, development of forecasts for longer periods and inclusion of meteorological variables produced by digital weather prediction models into the system.