Real-Time Carbon Emission Estimation for Industrial Users With Low RMSE Based on NILM and Evolutionary Algorithm

User-side carbon emission accounting is a means to sort out the amount and source of user carbon emissions. Real-time carbon estimation methods based on nonintrusive load monitoring (NILM) tend to miss the fact that there is not a negative correlation between equipment identification accuracy and carbon emission estimation error. Higher accuracy does not mean lower error. A two-stage training NILM network, LRMSE-ResNeSt, is proposed to reduce the error of real-time carbon emission monitoring results. It emphasizes the low root-mean-square error (RMSE) of carbon emissions' monitoring. The feature extraction network is first constructed using convolution and split-attention mechanism, and then, the model is trained using a two-stage training approach. In the first stage, backpropagation is used to update the network parameters for accurate device identification. In the second stage, the parameters of the fully connected layer are tuned using particle swarm optimization (PSO) to make the classifier more focused on the identification accuracy of devices with high carbon emissions. Finally, the proposed LRMSE-ResNeSt is validated using the industrial appliance identification dataset (IAID) industrial dataset. The experimental results show that the LRMSE-ResNeSt successfully reduces the RMSE of real-time carbon estimation by an average of 14.94%, which addresses the problem that the NILM method only focuses on the accuracy of the equipment identification but ignores the error of the carbon emission estimation results.