Plant data bases are filled with information on the relationships of state (output) and input variables. However, due to natural high multi-collinearities of the inputs and low signal to noise ratios for the outputs, modelers are challenged in developing cause and effect relationships using plant data. As a result, dynamic models are commonly developed from plant tests which incur costs and risks to the operation. Thus, the purpose of this work is to introduce a modeling approach that is capable of developing accurate cause and effect models from plant data. The proposed method is a special application of the Wiener block-oriented system and the unique and powerful attributes of this approach over existing techniques are demonstrated on a simulated continuous stirred tank reactor (CSTR) and real distillation column data.

Using the proposed method, the models are developed using high input correlation in training and then tested under conditions of low to zero input correlation in testing. Excellent performance of the models in testing validate the cause and effect ability of the modeling approach. The proposed method is compared with empirical methods like NARMAX which fail to predict well in testing, indicating their inability to model causation under high input correlation.