ANN Modeling and Optimization of Power output from Horizontal Axis Wind Turbine

Abstract

Integration of wind energy with energy with existing power sources has been restricted due to its intermittent and stochastic nature. Hence, there is a great need to develop an accurate and reliable site-specific prediction model. Forecasting of wind speed which is an important parameter affecting turbine power output, will help the wind energy industry in proper planning, scheduling and controlling. Artificial Neural Network (ANN) has proved its capability in mapping such complex non-linear inputoutput relations. The main objective of the wind energy industry, is to reduce the cost and increase the power generation by optimizing the controllable parameters affecting the turbine power output. The metaheuristic optimization algorithms, which are robust to dynamic changes are proved to be successful in solving such complex real-world problems. This research work has been carried out in three different phases namely wind power prediction, wind power optimization and wind speed forecasting (WSF).The data for this research work has been collected from the Supervisory Control and Data Acquisition System (SCADA) of 1.5 MW, pitch regulated, three bladed, horizontal axis wind turbine, located in a large wind farm present in central dry zone of Karnataka, India. In the present study, different conventional and ANN models have been used to predict the power output of a turbine. ANN models have been developed based on batch learning and Online Sequential Extreme Learning Machine (OSELM) algorithms, by considering carefully selected variables affecting power output, namely wind speed, wind direction, blade pitch angle, density and rotor speed. Maximizing the power output of the wind turbine by optimizing the only controllable parameter namely blade pitch angle has been achieved using three different metaheuristic optimization algorithms. A vihybrid ANN multistep WSF model, which is a combination of OSELM, Cuckoo Search (CS) and Optimized Variational Mode Decomposition (OVMD) method, hence named OVMD-CS-OSELM has been proposed in the present study. The performance of this hybrid model has been then compared with the benchmark models. From this study it has been found that, the models based on Extreme Learning Machine (ELM) converge extremely faster with better generalization performance and generate a compact network structure compared to Backpropagation learning. Out of the fifteen models based on batch learning, the fully optimized RBF model with ELM learning resulted in good performance with Root Mean Square Error (RMSE) value of 1.73%. The detailed study of OSELM algorithm showed a RMSE value of 1.96%, which is slightly higher than the fully optimized RBF model. However,for the present application due to the online nature of the wind data, OSELM algorithm is highly preferable. CS optimization algorithm is found to be suitable in optimizing the blade pitch angle of the turbine and accordingly the optimization of the power output, due to its fast convergence and a highest Mean relative PG value of 17.329%. In comparison with benchmark models, the proposed WSF model showed clear benefits of OSELM over ELM, OVMD over Emperical Mode Decomposition and CS over Partial Autocorrelation function for modeling, data pre-processing and input feature selection, with percentage improvements in Mean Absolute Percentage Error (MAPE) of 3.35%, 48.19% and 12.05% respectively for 1-step ahead forecasting. The proposed model has been validated using a standard database, which is from a meteorological station located in Portugal, thereby establishing its use in WSF. This research work thus proposes efficient models based on ANN for wind power prediction, optimization and WSF, which is useful in proper planning, integration and scheduling in the wind energy industry, thereby making it more competitive and a promising renewable energy source.