Gas turbines, an important energy conversion equipment, produce Nitrogen Oxides (NOx) emissions, endangering human health and forming air pollution. With the increasingly stringent NOx emission standards, it is more significant to ascertain NOx emission characteristics to reduce pollutant emissions. Establishing an emission prediction model is an effective way for real-time and accurate monitoring of the NOx discharge amount. Based on the multi-layer perceptron neural networks, an interpretable emission prediction model with a monitorable middle layer is designed to monitor NOx emission by taking the ambient parameters and boundary parameters as the network inputs. The outlet temperature of the compressor is selected as the monitorable measuring parameters of the middle layer. The emission prediction model is trained by historical operation data under different working conditions. According to the errors between the predicted values and measured values of the middle layer and output layer, the weights of the emission prediction model are optimized by the back-propagation algorithm, and the optimal NOx emission prediction model is established for gas turbines under the various working conditions. Furthermore, the mechanism of predicting NOx emission value is explained based on known parameter influence laws between the input layer, middle layer and output layer, which helps to reveal the main measurement parameters affecting NOx emission value, adjust the model parameters and obtain more accurate prediction results. Compared with the traditional emission monitoring methods, the emission prediction model has higher accuracy and faster calculation efficiency and can obtain believable NOx emission prediction results for various operating conditions of gas turbines.