Wind turbine power output very short-term forecast: A comparative study of data clustering techniques in a PSO-ANFIS model

Abstract

The emergence of new sites for wind energy exploration in South Africa requires an accurate prediction of the potential power output of a typical utility-scale wind turbine in such areas. However, careful selection of data clustering technique is very essential as it has a significant impact on the accuracy of the prediction. Adaptive neurofuzzy inference system (ANFIS), both in its standalone and hybrid form has been applied in offline and online forecast in wind energy studies, however, the effect of clustering techniques has not been reported despite its significance. Therefore, this study investigates the effect of the choice of clustering algorithm on the performance of a standalone ANFIS and ANFIS optimized with particle swarm optimization (PSO) technique using a synthetic wind turbine power output data of a potential site in the Eastern Cape, South Africa. In this study a wind resource map for the Eastern Cape province was developed. Also, autoregressive ANFIS models and their hybrids with PSO were developed. Each model was evaluated based on three clustering techniques (grid partitioning (GP), subtractive clustering (SC), and fuzzy-c-means (FCM)). The gross wind power of the model wind turbine was estimated from the wind speed data collected from the potential site at 10 min data resolution using Windographer software. The standalone and hybrid models were trained and tested with 70% and 30% of the dataset respectively. The performance of each clustering technique was compared for both standalone and PSO-ANFIS models using known statistical metrics. From our findings, ANFIS standalone model clustered with SC performed best among the standalone models with a root mean square error (RMSE) of 0.132, mean absolute percentage error (MAPE) of 30.94, a mean absolute deviation (MAD) of 0.077, relative mean bias error (rMBE) of 0.190 and variance accounted for (VAF) of 94.307. Also, PSO-ANFIS model clustered with SC technique performed the best among the three hybrid models with RMSE of 0.127, MAPE of 28.11, MAD of 0.078, rMBE of 0.190 and VAF of 94.311. The ANFIS-SC model recorded the lowest computational time of 30.23secs among the standalone models. However, the PSO-ANFIS-SC model recorded a computational time of 47.21secs. Based on our findings, a hybrid ANFIS model gives better forecast accuracy compared to the standalone model, though with a trade-off in the computational time. Since, the choice of clustering technique was observed to play a vital role in the forecast accuracy of standalone and hybrid models, this study recommends SC technique for ANFIS modeling at both standalone and hybrid models.