Traditional grid-connected photovoltaic converter controllers are highly sensitive to parameters and operating conditions, requiring time-consuming manual tuning of controller parameters and posing significant uncontrolling risks. Neural Network (NN) shows great potential as AI controller for PV converters. However, NNs are impractical for execution on low-cost MCUs, require relatively long training time and are low interpretability, leading to low acceptance in industry. These hinder converters from entering the “AI Gen”. This project proposes a Model Informed design for NN controller:(1) NN’s inputs and structure are chosen based on the converter model, minimizing the complexity of NN, making it can be implemented in a MCUs in real-time; (2) the training process is optimized based on the model information, accelerating the training convergence and thus, enhancing the suppression ability of control oscillation; (3) model knowledge is also used to understand the NN’s behaviors, making it to be well explained. Such an AI controller is implemented on DSP28377 in real-time to control a converter with 40kHz switching frequency. Both Hardware-in-the-loop and experimental results demonstrate the robust to parameters, operating conditions, and oscillation suppression. This idea had won the first prize in the national innovation competition certified by MIIT(Ministry of Industry and Information Technology).
Categories