Analysing historical data from turbines can identify faults and even predict potential failures. However, most articles focus solely on one single unit, overlooking the historical operational patterns of turbines from other units, which serve as important references for human experts. We propose ExpertAP, an approach that first learns the running patterns of turbines from historical data of different units and then performs anomaly prediction for the target turbine. This approach faces two key challenges: The scarcity of high-quality anomaly labels and the difficulty in using historical anomaly labels. Regarding the first challenge, we thus introduce a semi-supervised backbone which is pre-trained with the task of sequence reconstruction using data from multiple units and fine-tuned with the task of anomaly prediction using data from the target unit. We also propose a novel two-dimensional selection strategy: Filtering out anomaly labels in the dimension of the variable during pre-training and filtering out redundant normal time-series sequences in the dimension of time during fine-tuning. Regarding the second challenge, the anomaly labels are modelled as binary time series, which significantly differ from the sensor data generated from continuous sampling. Therefore, we design different embedding layers for the anomaly labels and sensor data. These layers are trained during fine-tuning to align the two types of input data in the latent space, allowing us to utilize historical anomaly labels as a basis for anomaly prediction. The proposed framework was tested on real data collected from the Turbine Supervision Instrumentation (TSI) system, showing promising test results.