稳压直流电源外文文献.doc

Neural Comput & Applic (2022) 18:93–103DOI 10.1007/s00521-007-0161-3O R I G I N A L A R T I C L EDesign of intelligent power controller for DC–DC convertersusing CMAC neural networkChun-Fei HsuReceived: 26 October 2022 / Accepted: 19 October 2022 / Published online: 8 November 2022Ó Springer-Verlag London Limited 2022Abstract DC–DC converters are the devices which canconvert a certain electrical voltage to another level ofelectrical voltage. They are very popularly used because ofthe high efficiency and small size. This paper proposes anintelligent power controller for the DC–DC converters viacerebella model articulation controller (CMAC) neuralnetwork approach. The proposed intelligent power con-troller is composed of a CMAC neural controller and arobust controller. The CMAC neural controller uses aCMAC neural network to online mimic an ideal controller,and the robust controller is designed to achieve L2trackingperformance with desired attenuation level. Finally, acomparison among a PI control, adaptive neural controland the proposed intelligent power control is made. Theexperimental results are provided to demonstrate the pro-posed intelligent power controller can cope with the inputvoltage and load resistance variations to ensure the stabilitywhile providing fast transient response and simplecomputation.Keywords Adaptive control Robust controlCMAC neural network DC–DC converter1 IntroductionThe DC–DC converters can convert a certain electricalvoltage to another level by switching action. The outputvoltage is controlled by adjusting the ON time of theswitching action, which in turn adjusts the width of aC.-F. Hsu (&)Department of Electrical Engineering, Chung Hua University,Hsinchu 300, Taiwan, ROCe-mail: fei@chu.edu.twvoltage pulse at the output. This is known as pulse-width-modulation (PWM) approach [1]. By varying the duty ratioof the PWM modulator, the DC–DC converter can convertone level of electrical voltage to the desired level. Formany years, the controller design for DC–DC convertershas been carried out through analog circuits, which limitedthem to PI controller structures [1, 2]. The selection of thecontroller parameters in the PI controller is a trade-offbetween robustness and transient response. In general, itinduces large overshoot in output voltage when the risetime of response is reduced.During the past decade, there have been many differentapproaches proposed for PWM switching control designbased on sliding-mode control [3, 4], fuzzy control [5, 6],and adaptive neural control [7–9] techniques. In the slid-ing-mode control design for DC–DC converters [3, 4], thecontrollers are easy to design and simple to implement.However, their performances generally depend on theworking point, thus these control parameters which want toensure proper behavior in all operating conditions are dif-ficult to design. In the fuzzy control design for DC–DCconverters [5, 6], the fuzzy rule should be pre-selectedthrough trial-and-error to achieve satisfy performances;however, this trial-and-error tuning procedure is time-consuming. In [7–9], though satisfactory regulator perfor-mance can be achieved by using the developed onlinetuning algorithm, the computation loading of these learningalgorithms is too heavy.The neural-network-based control technique has beenrepresented an alternative design method for various con-trol systems [10–14]. The successful key point is theapproximation ability of neural network, where the neuralnetwork can approximate an unknown system dynamics oran ideal controller after learning. Based on this property,the neural-network-based controllers have been developed12394Neural Comput & Applic (2022) 18:93–103to compensate for the effects of nonlinearities and systemuncertainties, so that the stability, convergence and robust-ness of the control system can be improved [9]. Recently,the cerebellar model articulation control (CMAC) neuralnetwork have been adopted widely for the control of com-plex dynamical systems owing to its fast learning property,good generalization capability, and simple computationcompared with the neural network [15–19]. The CMACneural network is classified as a non-fully connectedperception-like associative memory network with overlap-+Vi−QN1D1N2+Vx−D2LC+R Vo−ping receptive-fields [15]. The conventional CMAC neuralnetwork uses local constant binary receptive-field basisfunctions [20]. The disadvantage is that its output is con-stant within each quantized state and the derivativeinformation is not preserved. On the other hand, foracquiring the derivative information of input and outputPWMmodulatorFig. 1 The for。