AI-assisted design and control of next-generation power electronic converters for enhanced energy efficiency and power density in renewable energy systems

Abstract

Power electronic converter is widely used in many fields. The conversion efficiency is one of the most important specifications of power electronic converter, and it should be optimized because low efficiency would reduce production capacity and do harm to the device. However, with the complexity of algorithm and circuit structure increasing, it is always challenging to directly measure and calculate the efficiency of converter. Furthermore, traditional methods to evaluate the efficiency have numerous limitations, as many factors such as ambient temperature are not considered.

Due to the lack of the real-sampled data, in this project, the simulation model, a unidirectional 6.6 kW buck-boost converter, was first built up in Simulink to obtain a sufficient amount of simulation data. The simulation data is also tested by comparing the regulation in simulation data and the theoretical analysis.

Then, a BPNN model was built up in MATLAB, the input parameters and output efficiency of the simulation data are respectively set as the input layer and output layer of the model. Due to the lack of precise theoretical standards for setting the parameters of neural networks, parameters of the model need to be tested repeatedly. After analysis, when setting the number of neurons in hidden layer to 60, the model training could achieve relatively good speed and accuracy. Lastly, the model is tested by keeping some of the input parameters constant and choosing one or two parameters to be traversed to get the highest efficiency. The test result shows that the model could basically achieve accurate prediction of the converter efficiency.

The project also built up the hardware circuit board, including the test target – a circuit based on boost converter with higher power density, and an efficiency acquisition board. Furthermore, the ADC is partly configured on CCS by establishing the channel to connect the DSP and ADC, then the channel’s connectivity is tested with an C2000Ware example.

Lastly, the project researched on transfer learning to find ways to build up the connection between the real data and the simulation data.