Seite - 150 - in Adaptive and Intelligent Temperature Control of Microwave Heating Systems with Multiple Sources

5. ExperimentalResults Algorithm SIDMSE MeanSDofvalidation Time(percall) Nonlinear MISOEKF 0.0697 1.0698 0.00103s Linear MISORKF 0.1205 2.5691 0.00150s Linear MISORLS 0.1214 5.5467 0.00149s Linear MIMORKF 0.1068 4.1431 0.00148s Table5.1. Performancecomparisonof4grey-boxSIDalgorithms. the matrixAhave to be estimated, but in the MIMO approach the entirematrixisestimated. Moreestimationvariablesbringalarger estimationerrorandthereforedegradetheestimationaccuracy. According to the above results, in practical experiments, the MISO RKF algorithm is used for the linear MPC and the nonlinear MISO EKF isusedfor thenonlinearMPC. 5.2.2. Black-boxApproaches There are three algorithms introduced for the training of the NN, and theirsystemidentificationperformanceregardingtherealexperimen- tal data is shown in figures 5.15 and 5.16. As in the test of grey-box approaches, different algorithms were also tested for 20 times and the averagedparametersareshownintable 5.2. Above results approximately reflect the performance of the three dif- ferent learning in practice. SWBP has the minimum prediction MSE and a fast converging speed, but its validation error is the largest. More important, SWBP needs much more time compared with the other two algorithms. If the input or output dimensions increase, this execution time will become larger, which means it is not appropriate 150