The existing soft measurement methods of absolute gas emission generally do not consider the influence of the historical data of gas emission. To address this issue, a soft measurement model of gas emission based on the long short-term memory (LSTM) in deep learning is proposed. The time series of historical data of absolute gas emission and its related influencing factors are utilized for prediction. Due to the gradient problem, the LSTM model needs to pay special attention to control the learning rate to prevent the severe decreasing of prediction results. The LSTM cell structure is adjusted, and the softsign function is introduced to solve the gradient problem through its first derivative with relatively gentle changes. In this way, the network convergence is faster and less prone to saturation. In view of the existence of many hyperparameters in LSTM, the quantum-behaved particle swarm optimization ( QPSO) algorithm is used to optimize the soft measurement accuracy of absolute gas emission. And the kernel-principal component analysis is utilized to reduce the dimension of measurement indexes to accelerate the convergence speed of the model. Comparing the improved model with the initial model, the improved model has higher accuracy and efficiency. The root mean squared error, mean absolute percentage error and goodness of fit determinant are 0. 080, 0. 82% and 0. 988, respectively. Comparing the proposed model with ELM, PSO-SVM, PSO-BP and GRU models, the proposed model has smaller error and better measurement results than other models. Experimental results show that the proposed soft measurement model of gas emission has better performance.