Abstract: The prediction of loss on ignition of fly ash is the premise of its rational utilization.The conventional experimental method is expensive and often can not meet the field demand.In order to solve this problem, a convolution neural network based image recognition and loss on ignition prediction model of fly ash particles in electrostatic separation is proposed in this paper.The convolution neural network is improved by introducing Adam algorithm and dropout technology, Adam algorithm and dropout technology are introduced to improve the convolution neural network.The four characteristic parameters of gray mean value, gray variance, energy and entropy value of fly ash particle image are taken as the input, and the experimental value of fly ash loss on ignition is taken as the output.Two kinds of fly ash with high and low loss on ignition are collected from power plants in different regions for experimental research.Based on the experimental sample points, the loss on ignition prediction is carried out after learning and training.Considering the correlation and model training accuracy, the convolution neural network prediction model with learning rate of 0.010 is selected, and dropout technology is introduced.The numerical deviation between the convolution neural network model and the actual experimental data is 0.090 7, the numerical correlation of prediction is 0.975 0, and the prediction error is relatively small.The results show that over fitting phenomenon can be avoided effectively by dropout technology, compared with deep neural network, long-term and short-term memory network and BP neural network, convolution neural network model has better data feature extraction ability for fly ash image, and has a wide application prospect in the prediction of fly ash loss on ignition.