Building Soft Sensors using Artificial Intelligence: Use Case on Daily Solar Radiation

The concept of a soft sensor, a software replacement for unavailable, delayed or simply an expensive physical sensor, has received a lot of attention in the last decade. Its ability to estimate values of an observed phenomenon based on the expert knowledge defined in a form of theoretical or empirical models gives it a wide application and considerably lower costs. Furthermore, with the advent of artificial intelligence and data mining, soft sensors can be built purely by extracting correlations between different sensor data. However, both approaches can lead to either overfitting or overgeneralization. In this paper, both approaches are compared and combined in order to utilize the best of both worlds. Procedure for building soft sensors is defined and evaluated on a real-world use case for estimating daily solar radiation. The results show the best performance on the artificial neural network with RMSE of 1.454 and RRMSE of 3.42%.