In response to the issues such as the overly large scale of input parameters in the traditional LSTM data-driven model, which leads to an excessive computational burden, improper selection of hyperparameters, high frequency of turbine system failures, and high operation and maintenance costs, a turbine data-driven modeling approach based on PCA-POA-LSTM is proposed, and the turbine fault early warning is achieved by combining with the sliding window method. Firstly, the PCA dimensionality reduction technique is applied to reduce the dimension of the input data. Secondly, the POA parameter optimization method is adopted to select the optimal combination of hyperparameters. Then, the LSTM algorithm is utilized to predict the output parameters of the turbine. Finally, based on the prediction results of the PCA-POA-LSTM turbine data-driven model, the turbine faults are warned by combining with the sliding window method, and the alarm threshold is defined by the standard deviation within the window, thus conquering the difficulty of turbine fault early warning. The results indicate that the turbine data-driven modeling based on PCA-POA-LSTM achieves a relatively high accuracy, with the average absolute percentage error all below 0.396, the average absolute error all below 0.809, and the average root mean square error all below 1.387. Moreover, the fault early warning method can issue a fault early warning signal at least 173 monitoring points in advance, achieving the purpose of turbine fault early warning and providing theoretical basis and technical support for the future development of turbine health management.