FPGA Implementation of an Efficient Neural Network Model for Maximum Power Point Tracking

In this paper, field programmable gate array (FPGA) implementation of a compact, efficient fully-connected neural network (FCNN) model is presented which can track maximum power point of a radio frequency (RF) based energy harvesting system. The hyperparameters of the FCNN model have been tuned to obtain optimal accuracy rate of nearly 78% and have been optimized to ensure compact design solution in miniaturized CMOS circuits. A two-hidden layer FCNN model was trained to optimize the binary cross entropy loss function. To prevent the FCNN model from overfitting, ADAM optimizer was used during the training phase. Weights and biases extracted from the best learned FCNN model were implemented into digital hardware. The application-specific trained neural network model embedded in hardware platform requires minimal resources which are significantly less compared to commercially available hardware accelerators. In future, the FCNN model can be integrated into a standard system-on-chip platform for developing RF energy harvester.