The invention discloses a trend analysis based transformer latent fault early warning method. The method comprises the steps that on the basis of a non-parametric regression method, historical data issubjected to smoothing process, and by means of detection and evaluation indexes of abnormal values such as the accuracy rate, recall rate and bandwidth, an optimal smoothing factor and upper and lower limit sequence data corresponding to the optimal smoothing factor are obtained; by taking the upper and lower limit sequence data as the historical data, a gas concentration self-adaptive prediction model is built, and the upper and lower limit gas concentration data in the future time period is predicted; by means of comparison of actual detection data and predicted upper and lower limit data,an early warning strategy is determined. According to the method, the problems such as missing report and misinformation which exist in a fixed threshold value are avoided, the field application requirements can be met, the early warning reliability of transformer latent faults is improved, more reliable reference is provided for maintenance work of a transformer, and safe and stable running of the transformer and an electric power system is guaranteed.