A big data analytical framework for analyzing solar energy receptors using evolutionary computing approach

Data science has been empowered with the emerging concept of big data enabling data scalability in many ways. Effective prediction systems for complex analytical problems dealing with big data can be created using evolutionary computing, associate feature selection and reduction techniques. In the current work, we put forward a big data analytical scheme to analyze solar energy receptors based on a set of features. Correct estimation of pressure loss coefficients (PLC) greatly improves the design of a solar collector. Evaluation of PLC is a time and resource consuming process as the flow rate and Reynolds number changes at every junction. Moreover, a suitable and appropriate algebraic expression is not yet defined in the laminar region of flow for approximation of the complex relationship among different geometrical features and flow variables. The overall heat gain of the solar receptor is dependent upon flow rates and flow distribution in risers. Also, the local disturbances during the flow division and combining process from manifold to risers affects the performance of the solar collector. Owing to these reasons, mostly they are calculated using experiments, primarily due to the complexity involved. The proposed big data framework involves acquiring huge feature sets at each point along the flow of thermal fluid. The data is experimentally acquired in a set of around forty features for large number of Reynolds number and discharge ratio variations. Reynolds number varies from 200 to 15,000 while discharge ratio variation is in the range of 0–1. Feature reduction in the big data set is done by calculating the relevancy score using ReliefF algorithm that extracts the most relevant features. Later, the framework employs a suitably selected optimal ANN architecture of layers, neurons and activation functions. The selected topology is trained using reduced features sets using Levenberg–Marquardt backpropagation algorithm. Test and validation results bespeaks the efficacy of the proposed strategy and indicate that future PLC values can be forecasted close to experimental data. The relative percent error is around 10% of the experimental data set and is found better than computational fluid dynamics based approaches in terms of memory and processing time.