Improvement and application of MobileNetV3-small model in the field of smart grid

In this paper, we consider the problem that artificial neural networks consume much more computational resources than needed for real-world scenarios during their deployment on the power grid. Artificial intelligence has a huge number of models in fields such as computer vision, but the cost required for model training and deployment is very high. For the reason that power grid task scenarios are very specific, this paper makes a targeted adaptation and improvement of the MobileNetV3-small model. The new model we propose is MobileNetV3 for Smart Grid (MSG), which further reduces the required computing power and time while the inference accuracy is almost unchanged and speeds up the inference speed. Compared with the original h-swish activation function in MobileNetV3-small, the average computation time of the proposed h-tanh activation function in MSG has been tested to be 7.3% shorter. Under the same hardware conditions, the inference speed of MSG can be improved by 7.3%-11.2%, depending on the model hyperparameter settings. At the same time, the inference accuracy is improved by 10%-15% compared with that of Support Vector Machine (SVM) in different scenarios of image recognition applications.