The present invention relates to the technical field of water level prediction of ship nuclear power voltage regulators, and particularly relates to a voltage regulator water level prediction method based on a cost-sensitive LSTM cyclic neural network. The voltage regulator water level prediction method based on the cost-sensitive LSTM cyclic neural network comprises the following steps: S1, selecting p parameters (as shown in the specification) with a relatively high degree of coupling with a voltage regulator water level as input parameters; S2, constructing a LSTM-based voltage regulator water level prediction model and a framework thereof; and S3, using a BPTT algorithm to train and optimize the water level prediction model constructed in step S2. According to the method provided by the present invention, the LSTM model can better approximate the true value of the water level than the SVR model and the BP neural network model, and has stronger learning ability and prediction ability, and the cost-sensitive LSTM model has better precision and faster convergence.