Effectiveness of evolutionary-tuned neurofuzzy inference system in predicting wind turbine gearbox oil temperature

Monitoring the gearbox oil temperature of wind turbines has been one of the most effective ways of increasing their reliability, and reducing maintenance and operational costs. The temperature variation is affected by several factors, including meteorological variables. Unfortunately, this is less considered in the condition monitoring of wind turbine gearboxes. Also, while the neurofuzzy inference model offers a high prospect in improving the understanding of wind turbine gearbox oil temperature behaviour, tuning its hyperparameters for such an application has been less explored. Hence, this study investigates the effectiveness and efficiency of predicting wind turbine gearbox oil temperature using an adaptive neurofuzzy inference system optimized with population-based evolutionary techniques (the genetic algorithm and particle swarm optimization). Relevant meteorological variables were selected from a case study wind farm in South Africa and used as model inputs. The hyper-parameter tuning of the neurofuzzy inference system model was carried out using population-based techniques before training the models. The performances of these hybrid models were compared based on established statistical performance metrics. The genetic algorithm hybrid model outperformed the particle swarm optimization hybrid on the test dataset with a root mean square error of 3.79, mean absolute deviation of 2.93, mean absolute percentage error of 5.9, and relative mean bias error of 1.74. The genetic algorithm-based hybrid model showed higher accuracy and reliability for wind turbine gearbox oil temperature prediction in the case study wind farm and offers potential use for neighbouring wind farms in the same renewable energy development zone.