Development of a deep learning and feature extraction - based forecasti̇ng method for short - term wind speed

In this thesis, a signal decomposition and deep learning based method has been developed for short term wind speed forecasting. Only wind speed was used at the input of the developed network, and the wind speed data were decomposed into their sub-components using the Variational Mode Decomposition (VMA) method before being applied to the proposed model. The obtained intrinsic mode function components were converted into a matrix and given as input to the proposed model, and the forecasting process was carried out with a hybrid architecture consisting of Convolutional Neural Network (CNN) and Bidirectional Long Short Term Memory (Bidirectional LSTM) layers. While the CNN layers extracted structural features from the decomposed signal, the Bidirectional LSTM layers modeled temporal dependencies bidirectionally, increasing forecasting accuracy. The proposed model was tested using hourly wind speed data from 2019 and compared with the CNN - LSTM (Long Short Term Memory) and CNN - GRU (Gated Recurrent Unit) models across all forecasting horizons. The performance of the proposed model was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Correlation Coefficient (R), and Symmetric Mean Absolute Percentage Error (SMAPE) metrics. The proposed CNN - Bidirectional LSTM model achieved an R value of 0.9680, an RMSE value of 0.8417, an MAE value of 0.6263, and a SMAPE value of 0.1280 for the 1 hour-ahead forecast; an R value of 0.9418, an RMSE value of 1.1479, an MAE value of 0.8654, and a SMAPE value of 0.1730 for the 2 hour ahead forecast; and an R value of 0.9165, an RMSE value of 1.3294, an MAE value of 1.0098, and a SMAPE value of 0.1997 for the 3 hour ahead forecast. These results show that the proposed model demonstrates high accuracy in short term forecasting. In addition, the proposed model was compared with the CNN - LSTM and CNN - GRU architectures and achieved lower error and higher accuracy in all forecasting horizons, being identified as the most successful method. The obtained results show that the proposed model exhibits competitive performance and can be effectively used in smart grid operations.