In response to the problems of low diagnostic accuracy and inconsistent characteristic parameter standards in current power transformer fault diagnosis methods, based on the analysis of power transformer faults, a fault diagnosis method of power transformer based on bidirectional short-term memory (BiLSTM) network and improved whale optimization algorithm is proposed. The hybrid strategy (weight and convergence factor optimization, bat algorithm and Levy flight strategy) is introduced to optimize the whale optimization algorithm, and the optimized whale optimization algorithm is used to find the optimal parameters of the bidirectional short-term memory network to establish the power transformer fault diagnosis model. The superiority of this method is verified through comparative analysis between numerical examples and conventional methods. The results show that compared to conventional methods, the proposed fault diagnosis method has higher fault diagnosis accuracy and the best practical application effect, the fault diagnosis accuracy has been improved by 10.42% and 7.85%, providing a new approach for power transformer fault diagnosis.