SAMSE 2018 IOP Conf. Series: Materials Science and Engineering 490 (2019) 062027
IOP Publishing doi:10.1088/1757-899X/490/6/062027
v n
2 l y y ¦ between the experimental value l y and the predicted value of the model of the effluent COD as the objective function, through the particle swarm optimization (PSO) (Kennedy et al., 1995), based on the validation sample set. 2.2.3.Model Parameter Adaptive Correction 1 ˆ ( ) l l In order to track the dynamic changes of the production process and maintain the prediction performance of the soft-sensor model in real time, an online calibration strategy (OCS) has been designed to iteratively update the soft-sensor model parameters in an adaptive manner. The basic idea of OCS is that if the established soft-sensor model is applied to the prediction of COD for a new sample individual, when the deviation between the experimental value new y of the new sample individual and the predicted value ˆ new y of the model exceeds the set error limit maxe , namely: ˆ | | new new y y maxe ! (10) To Initiate the iterative update of the soft-sensor model parameters. The specific method is as follows: firstly, the sample individuals with the largest fitting deviation are retrieved from the training sample set and deleted; then, the sample individuals with the highest ranking in the monitoring sample set are transferred into the training sample set; next, the vacancy of the validation sample set is filled, namely, the sampled individuals with the highest ranking among the accumulated predicted sample individuals are transferred into the validation sample set; and finally, the soft-sensor model is re-established based on the newly formed training sample set and the validation sample set, which is namely the OCS - PCA-PSO-LSSVM model.
X
new X
new y
new T
T
y
ˆ ! new new maxe y y
{ , } T y
{ , } V J
{ , } X y
ˆ y
ˆ new y
Fig 2. Structure of the OCS-PCA-PSO-LSSVM soft sensor model With this, the implementation flow of the OCS-PCA-PSO-LSSVM method is shown in Fig. 2. Firstly, the PCA pre-processing of T 1 2 [ ] n p p u X = x ,x , ,x was performed and the pivot element matrix T 1 2 [ , , , ] n d d u T t t t was obtained after the number d of pivot elements had been selected; then, based on n d u T and the output variable matrix T 1 1 2 [ , , , ] n n y y y u y of the effluent COD, the nonlinear mapping relationship between them was established with the LSSVM method, while the values of
5
Made with FlippingBook Digital Publishing Software