LEE- Enerji Bilim ve Teknoloji Lisansüstü Programı

Bu topluluk için Kalıcı Uri

http://hdl.handle.net/11527/19269

Gözat

Konu "artificial neural networks" ile LEE- Enerji Bilim ve Teknoloji Lisansüstü Programı'a göz atma
  • Öge
    Predicting electricity production in a wind farm using artificial neural networks
    (Graduate School, 2023-01-24) Akyol, Esma ; Barutçu, Burak ; 301191011 ; Energy Science and Technology
    In this thesis, it is aimed to make an estimation of electricity generation with the help of artificial neural networks using the ten-minute average wind speed (m/s) and average power (kW) data of 2015 recorded from a turbine belonging to a wind farm in Çanakkale. For this purpose, four different models were established. The MATLAB programming language was used during model building and operation. Of the 52108 data in the data set, 70% (36476) was used for training, 15% (7816) for testing, and the remaining 15% (7816) to be used in validation and used the same in all models. The lag value was taken as 10, and since the data was 10 minutes, the estimations were made for 100 minutes beyond. In the first model, wind speeds at time t in the input layer and power generation at time t in the output layer are given. The model was run by increasing the number of neurons in the hidden layer in order to make comparisons and check whether there is improvement. 5 neurons in the first two trials, 10 neurons in the third trial, and 15 neurons in the fourth trial; the model was trained by running it twice in trials with 5 neurons and three times in trials with 10 and 15 neurons. The activation functions used for each trial, respectively; logsig-tansig, tansig-tansig, logsig-tansig and logsig-tansig. 𝑅2 values were observed as 0.978577, 0.978491, 0.978658, 0.978647, respectively. It was concluded that number of neurons is an important parameter where if too much neurons are used for training, the error results will increase. In the second model, the wind speeds at time t and t-1 are given as input to the system, while power generation values at time t are used as output. Since there are two different inputs in the input layer, the number of neurons in the hidden layer and the number of training the model have been increased. The model was trained so that it was 15 in the first trial, 20 in the second, and 25 in the third and fourth trials, and the system was run five times. Activation functions were determined as logsig-tansig, logsig-tansig, logsig-tansig and logsig-logsig, respectively. 𝑅2 values were obtained as 0.967898, 0.968066, 0.968068, 0.642207, respectively. It was concluded that the error parameter difference in the experiments of the model working with 25 neurons occurs when the logsig activation function is used instead of tansig in the output layer. In the third model, wind speeds at time t and power generation data at time t-1 are used in the input layer, while power generation at time t is given to the system in the output layer. The model was trained by running the system five times, with the number of neurons in the hidden layer increasing to 25; 15 in the first and second trials, 20 in the xxii third. Activation functions are determined as logsig-logsig, logsig-tansig, logsig-tansig, respectively. The 𝑅2 values are 0.996571, 0.996633, 0.996716, respectively. In the fourth model, power generation values at time t-1 and t-2 are given to the input layer, while power generation values at time t are used in the output layer. The model was trained by running the system five times, with the number of neurons in the hidden layer increasing to 25; 15 in the first trial, 20 in the second, and 25 in the third. Activation functions were determined as logsig-tansig, logsig-logsig, logsig-tansig, respectively. 𝑅2 values were obtained as 0.984357, 0.652384, 0.984398, respectively. In the study, it is also considered to use different meteorological parameters as input data that can be used to predict the electricity to be produced from a wind turbine. However, it has been investigated that the effects of parameters such as temperature and relative humidity on the values in turbine production are neither positive nor relevant. On top of that, it is aimed to obtain predictions by changing the model network. In the most basic step, the wind speeds at the same time were used, and then the wind speed data of the previous time step were also used. Likewise, the effect of the power data of the previous time steps on estimating the power generation of the next time step was examined. The importance of the activation functions according to the data type has been pointed out, and logistic sigmoid and hyperbolic tangent functions are used in the functions used in the transition from the input layer to the hidden layer and from the hidden layer to the output layer. In addition, it has been investigated that changing the number of neurons in the hidden layer is a factor that has an effect on the training of the system. It is important to find the limit number as it will cause it to happen if too many neurons are used. The number of times the system would run and train varied between 2 and 5 depending on the number of inputs. However, adding too many hidden layers and neurons can also lead to overfitting, where the network becomes too specialized to the training data and performs poorly on new, unseen data.