The invention relates to a similarity-based long-term prediction method and system for engine gas path parameters. The method comprises the following steps of: calculating time sequences of point-by-point distance features between a target track and history tracks; assessing a statistical distance between the target track and each history track according to the obtained time sequences of the point-by-point distance features; aiming at single feature elements on each prediction time point, enabling each history track sample to generate assumed Gaussian function-form probability density estimation by utilizing the obtained statistical distances and history track samples, so as to form an assumed Gaussian unit set; and aggregating the obtained assumed Gaussian unit set through a descending-order aggregation method so as to obtain a Gaussian hybrid model of a target feature. Compared with auto-regression moving average prediction methods, counter-propagation neural network prediction methods and traditional similarity-based prediction methods, the method provided by the invention has higher prediction precision.