The deep reinforcement learning (DRL) is widely utilized in various dynamic systems to search for optimal control strategies through interactions with the system environment. However, the training performance of the DRL method for nonlinear hysteresis systems subject to sensor disturbances (SDs) is limited by the instability of the training data distribution. This study proposes a temporal feature enhancement-based deep deterministic probability prediction (TFE-DDPP) scheme to enhance the training performance of the DRL agent for the magnetorheological fluid-based semi-active suspension (MRF-SAS) system with SDs. First, the temporal signals acquired from sensors are transformed into feature matrices by using Gram’s corner field method. Subsequently, these feature matrices are further extracted by using the convolutional neural networks (CNNs) to enhance the feature representation of the temporal signals. Moreover, the extracted feature matrices and the temporal signals are integrated by using the multi-head attention method to avoid the loss of important features from the raw signals. In addition, the moving standard deviation function is introduced to address the instability in the distribution caused by SDs. Moreover, the dynamic model of the MRF-SAS system is constructed in MATLAB/SIMULINK, while the DDPP agent is developed in PYTHON to conduct the co-simulation. Experimental results show that the proposed TFE-DDPP scheme improves the training performance of the DRL agent for the MRF-SAS system with SDs.