Power prediction holds significance within the Energy Management System (EMS) of Hybrid Electric Vehicles (HEVs). While deep learning methods offer benefits such as excellent fitting ability and the capacity to approximate complex functions, their application to energy management systems with numerous control parameters can lead to complexity and reduced operational efficiency. Therefore, dimensionality reduction of control parameters is crucial for computational efficiency. This paper compares four dimensionality reduction methods—L2 Norm, Pearson Correlation, Principal Component Analysis (PCA), and Spatial Predictor Envelope (SPE)—and their integration with the Bi-LSTM network to evaluate power prediction performance. The results demonstrate that the SPE method exhibits the best overall performance. Specifically, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of SPE are 6.64% and 2.97% higher than PCA, respectively. However, the training time is shorter than PCA by 2.81 seconds. These results collectively highlight the superior and effective contribution of the SPE method, emphasizing its efficiency in enhancing the performance of deep learning models.