A hybrid machine learning using Mamdani type fuzzy inference system (FIS) for solar power prediction

This paper empirically demonstrates that a hybrid or aggregated machine learning using Mamdani type fuzzy inference system (FIS) for solar power prediction can deliver improved prediction accuracy that outperforms those of single machine learning technique. However, for aggregation to be effective, the individual machine learning techniques must be as accurate as possible. Therefore, this paper is presented in two major phases. In the first phase, the selection procedure of those individual machine learning techniques and the adopted strategies to make them as accurate as possible are briefly explained. According to the experimental results the most potential three machine learning techniques namely Least Median Square (LMS), Multilayer Perceptron (MLP) and Support Vector Machine (SVM) were selected for the components of the hybrid prediction method. After applying each of the improvement strategies on the machine learning techniques, six hours ahead solar power predictions were performed and justified with the error validation metrics. The prediction outcome of each and every step was compared with the prediction outcome of the previous steps to ensure the continual improvement of the prediction accuracy of those individuals. All the prediction accuracy results executed by the individuals in the first phase are summarized which clearly reveals a significant contribution of this paper. It is found from the summarized results that those individual machine learning techniques reach at the best of their accuracy level by the combined effect of applying feature selection and parameter optimization on them. In the second phase, the best possible individual accurate predictions from LMS, MLP and SVM are passed as inputs towards the Mamdani FIS to accomplish the task to non-linearly integrate them. From the empirical results it is obvious that the final aggregated prediction delivers better accuracy than any of the individuals in terms of the error validation metrics Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Mean Absolute Scaled Error (MASE). All the experiments in this paper are performed using a reliable and real life solar radiation dataset which is collected over five years with hourly resolution from 2006 to 2010 for Rockhampton, Australia. The experimental results are also illustrated in graphical and three dimensional surface views for the better and easier understanding.