One-hour ahead wind speed forecasting using deep learning approach


Stochastic Environmental Research and Risk Assessment, vol.36, no.12, pp.4311-4335, 2022 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 36 Issue: 12
  • Publication Date: 2022
  • Doi Number: 10.1007/s00477-022-02265-4
  • Journal Name: Stochastic Environmental Research and Risk Assessment
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Compendex, Environment Index, Geobase, Index Islamicus, Pollution Abstracts, zbMATH, Civil Engineering Abstracts
  • Page Numbers: pp.4311-4335
  • Keywords: Deep learning, Wind speed, One-hour ahead prediction, ANFIS-FCM, ANFIS-GP, ANFIS-SC, LSTM
  • Ankara Yıldırım Beyazıt University Affiliated: Yes


© 2022, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.Accurate forecasting of wind speed (WS) data plays a crucial role in planning and operating wind power generation. Nowadays, the importance of WS predictions overgrows with the increased integration of wind energy into the electricity market. This work proposes machine learning algorithms to forecast a one-hour ahead short-term WS. Forecasting models were developed based on past time-series wind speeds to estimate the future values. Adaptive Neuro-Fuzzy Inference System (ANFIS) with Fuzzy c-means, ANFIS with Grid Partition, ANFIS with Subtractive Clustering and Long Short-Term Memory (LSTM) neural network were developed for this purpose. Three measurement stations in the Marmara and Mediterranean Regions of Turkey were selected as the study locations. According to the hourly WS prediction, the LSTM neural network based on the deep learning approach gave the best result in all stations and among all models applied. Mean Absolute Error values in the testing process were obtained to be 0.8638, 0.9603 and 0.5977 m/s, and Root Mean Square Error values were found to be 1.2193, 1.2573 and 0.7531 m/s from the LSTM neural network model for measuring stations MS1, MS2, and MS3, respectively. In addition, the analyzes revealed that the best correlation coefficient (R) results among the algorithms in the test processes were obtained to be 0.9498, 0.9147, and 0.8897 for the MS1, MS2, and MS3 measurement stations, respectively. In this regard, it is shown that the LSTM method gave high sensitive results and mainly provided greater performance than the ANFIS models for one hour-ahead WS estimations.