Wind speed forecasting using neural networks and effects of frictional power loss in wind turbines

Abstract

This study consists of wind potential forecasting based on artificial neural networks (ANN) in order to quantify the available energy and to allow an optimal management of the transition between different sources of energy. It also enabled the quantifica- tion of power losses of the wind turbine gearbox that can improve its performance and therefore the power generation. Five ANN architectures with the wind data collected at Bapouh in Cameroon from 19th November 2016 to 31st December 2017 after every 10 minutes at a height of 70 meters are used for application. The autoregressive integrated moving average (ARIMA) model is used as the main comparison model. Firstly, we analyze the performance of the six mostly used forecasting models for multi-step ahead wind speed forecasting. The best obtained results are used to estimate the expected electrical power generation by E-82 2000 kW wind turbine. These results showed that the non-linear autoregressive exogenous neural network (NARXNN) and ARIMA models provide better performance than the other models. It was shown that the adaptive neuro-fuzzy inference system (ANFIS) model is the worst model compared to the others for multi-step ahead wind speed forecasting. The best forecasting performances are obtained at 1-step ahead, while the worst are obtained at 3-step ahead. Also, it has been shown that the prediction accuracy of the expected wind turbine power generation depends closely on the wind speed forecasting accuracy. Subsequently, due to the unstable character of the relationships between the weather variables dependent on the wind speed, a novel hybrid model combining ANN and ARIMA models is proposed to improve the performance of the wind speed forecasting obtained from the previous traditional ANN models. This proposed hybrid model is based on the improvement of the relationship between two weather variables. Thus, a new approach to transform actual data before using them for ANN training is proposed. The experimental results indicate that the proposed data transformation strategy is appropriate in strengthening the relationship between two variables and decrease the seasonal variation. Moreover, in terms of forecasting accuracy, the proposed hybrid model outperforms other comparable models. Also, we analyse the influence of the ANN input variables disposition (IVD) on its training and forecasting performances. The obtained results for a static model and a dynamic model of ANN showed that their performances change differently with their IVD. This proves that it is necessary to include the optimal IVD in ANN optimization. Thus, a new approach to ANN optimization is proposed by introducing the IVD into back-propagation algorithm used for ANN training. This approach is validated on three traditional ANN models. The obtained results reveal that each proposed model is superior to its traditional model in terms of wind speed forecasting accuracy. Finally, due to the complexity of setting up a model for predicting mechanical losses in the wind turbines, the SKF model associated with the ANN is proposed for real-time power losses in the rolling bearings of the wind turbine gearbox modeling and prediction. The SKF model is used to determine the historical values of power losses. The back-bropagation neural network (BPNN) model is used for the historical values modeling and predicting the desired values. The achieved results revealed that the bearing power loss is highly influenced by the wind turbine operating parameters, capacity, and oil. The difference between actual and neural network predicted bearing power loss values under real-time operating parameters showed the effectiveness of the proposed approach.