A new hybrid deep-learning model with hyperparameter optimization for short-term wind power generation forecasting

The present research paper presents a new hybrid deep-learning model which aims to enhance the precision of multi-step, short-term wind power generation forecasting. The proposed hybrid model integrates two distinct deep learning models i.e. Bi-LSTM (bidirectional long-short-term memory) network and 1D-CNN (one-dimensional convolutional neural network). The fine-grained features from the original input record are automatically extracted by a convolutional layer in a 1D-CNN. The Bi-LSTM layers are used to retain crucial information, allowing it to remember and process data over a longer time. To further improve predicting accuracy, two distinct methods i.e. discrete wavelet transformation (DWT) and adaptive random search optimization (ARSO) have been applied to denoise the time-series input data of wind power and optimize model’s hyperparameters, respectively. The proposed prediction model is tested on SCADA datasets at the Yalova Wind farm, covering the period from January 2018 to December 2018, which considers three input features: wind speed, wind direction, theoretical power curve, and whereas wind power output is treated as the target variable. The effectiveness of the model's performance is examined using 10-minute data from the Yalova Wind Farm, focusing on forecasting horizons of 30, 40, and 60 minutes. The proposed model is also compared with benchmarking models such as Bi-LSTM, 1D-CNN, LSTM, gated recurrent unit (GRU), and hybrid 1D-CNN+Bi-LSTM, is to analyze the model's robustness. The output evaluation metrics, i.e. mean squared error (MSE), root-mean-square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ), reveal that the proposed hybrid model is found superior in terms of overall forecasting precision across all three forecasting horizons. For the 30-minute forecasting horizon, the proposed model outperforms other competing models with superior average values for all the error metrics i.e. MSE = 0.0109, R 2 = 0.9206, RMSE =0.1044, and MAE = 0.0656. Moreover, it is capable to achieve mean values of output error metrics i.e. MSE = 0.0135, R 2 = 0.9016, RMSE = 0.1163, and MAE = 0.0710 for 40-minute forecasting horizon, and MSE of 0.0187, R 2 of 0.8751, RMSE of 0.1368 and MAE of 0.0812 for 60-minute forecast horizon. ✓ The proposed hybrid model is compared with several other benchmarking forecasting models, such as GRU, 1D-CNN, Bi-LSTM, LSTM, and hybrid 1D-CNN+Bi-LSTM, to test its effectiveness. ✓ Its performance is further evaluated and verified for three forecasting horizons of 30-Min, 40-Min, and 60-Min, which offers deeper insights into the model's effectiveness for multistep forecasting applications. ✓ The output forecasting precision of the proposed hybrid deep-learning model is examined on four distinct performance metrics i.e., MSE, RMSE, R 2 , and MAE, demonstrating its model's practical relevance with their significant improvements. ✓ An extensive statistical analysis with Boxplots and Diebold-Mariano test results w.r.t. to other competing models are presented to validate the robustness of the proposed hybrid model.