Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Application and deep learning algorithms for the short-term and medium-term wind speed prediction

Kourtidis Dimitrios

Full record

---

Summary

Wind speed forecasting plays a pivotal role in various industries, particularly in the realms of environmental management and investments. Accurate predictions of wind speed are crucial for optimizing the performance of wind energy systems, aiding in efficient power generation and grid integration. Machine learning methods for wind speed forecasting leverage historical meteorological data and advanced algorithms to enhance prediction accuracy. Supervised and unsupervised learning techniques, including regression and neural networks, analyze past weather patterns to identify relationships and patterns for predicting future wind speeds. A big challenge in machine learning lies in the size and quality of the available datasets. When dealing with small datasets, models may struggle to generalize effectively, potentially leading to overfitting and limited predictive capabilities. Due to the small of the used dataset in this study, one of the primary goals is to increase its size using signal augmentation techniques. This was necessary for increasing the efficiency of wind speed forecasting which counts as the second goal of this thesis. Τhe generation of the new data was done in the frequency domain according to the Adjusted Amplitude Fourier Transform (AAFT) algorithm. Then, the goal was to develop and apply advanced deep learning for wind speed forecasting. The two models were a stacked autoencoder (SAE) and stacked independently recurrent autoencoder (SIRAE). The innovation in the present work lies in the design, implementation and application of SAE and SIRAE on the basis of signal based augmentation by means of AAFT. We first apply the AAFT algorithm to increase the size of the used dataset and we then divide it by month due to the seasonality of the wind speed,then we create subsequences, of different sizes, of the data that will be taken into account for the current forecast and then we train the models. The aim of this work is to create models with very high success rates and robustness and to be compared with the results from literature as well as the most suitable data subsequence size. The comparison between these two models made taking into account the MSE and RMSE of the forecasts. Both models showed high performance, so to draw clearer conclusions the RMSE of the two models was averaged.

Available Files

Services