Deep Learning-based Adaptive Control of DC-DC Converters Under Variable Load Conditions

Abstract

This paper presents a deep learning (DL)-based adaptive control framework to be implemented for the buck-type DC-DC converter under a variable load scenario. By relying on real-time electrical properties, namely input voltage, output voltage, inductor current, output current, and past duty cycle as well as voltage error and its evolution (variation), the controller predicts a duty cycle optimal for pulse width modulation (PWM) switching. State-space modelling is used to describe converter dynamics, the behaviour of inductor current and capacitor voltage response as well as load resistance estimation and voltage tracking error formulation. The function of the trained neural network (NN) translates the measured operating states into a necessary duty ratio, and saturation limits keep the switching operation in safe bounds by filling pre-defined gates of the duty cycle. The purpose function minimises voltage tracking error and suppresses duty cycle variation beyond tolerance limits for better transient stability and switching stress. The model will utilise the design for input source 24 V, reference output as 12 V, rated power of 120 W, switching frequency of 50 kHz, an inductor of size at 220 μH & a capacitor as output size at 470 μF. A simulation-orientated analysis of carrier comparison, gate pulse generation, inductor current response, output voltage ripple and performance variation at various load conditions. It can be found from the comparisons that the suitable AC voltage adaptive duty correction is well realised, and output voltage deviation decreases obviously by using a DL-based controller compared with traditional fixed-parameter control strategies, while stable operation under sudden load disturbance is guaranteed.