Neural network approach based on convex incremental learning machine for prediction of diffuse solar radiation

This paper introduces an alternative way of modeling the solar diffuse radiation based on extreme learning machine methods, which are gaining a growing interest in the scientific and research community nowadays. Several models are built that employ the classic, incremental and convex incremental extreme learning algorithm, and are compared to each other, as well as to other available models, in order to evaluate their approximation capability and accuracy. Along with the models, a few important features of the learning algorithms are discussed and alternative solutions are offered for some learning steps. Namely, the conducted research has clearly showed that the random selection of the hidden layer parameters significantly influences the approximation capacity of the classic extreme learning machine. The paper offers a simple solution to the problem. In addition, the research has confirmed that the incremental extreme learning machine indeed does not achieve the smallest possible approximation error due to the fact that the output parameters of the hidden nodes are not readjusted after the addition of each new hidden node. The paper also offers a simple solution to this problem. Finally, the convex incremental extreme learning machine tends to solve the accuracy problem of the incremental extreme learning machine. However, it still achieves smaller accuracy than the proposed solution in this paper. Nevertheless, the simulation results within this research show clearly that the extreme learning machine methods indeed possess the attributes of extreme simplicity, extremely good approximation performance, and extremely fast computation.