<p>Large forging presses are commonly employed in heavy-duty forging production lines within the mechanical manufacturing sector, playing a pivotal role in advancing forging technologies across industrial domains such as automotive manufacturing, shipbuilding, and aerospace. Typically, a large forging press consists of multiple components with intricate and highly coupled relationships. Consequently, the early identification and accurate early warning of fault signals in large forging presses become exceedingly challenging. Furthermore, due to the concealed, highly nonlinear, and stochastic characteristics of the relationships between components, traditional single-model state prediction methods often suffer from overfitting or underfitting, ultimately resulting in insufficient generalization capability and subpar prediction performance. To address these issues, this paper proposes a multi-scale Autoregressive-Support Vector Regression-Long Short-Term Memory (AR-SVR-LSTM) multi-model ensemble prediction approach. This method leverages the AR model, SVR model, and LSTM neural network model to predict the states of critical components in large forging presses. By capitalizing on the strengths of each model, a hybrid prediction model with weight constraints is introduced. Finally, this study validates the proposed ensemble model using monitoring signal data of brake oil pressure from an 80MN electric screw press production line at a certain enterprise as a case study. The research findings indicate that the accuracy of the proposed model significantly falls short of the industry standards for such models in industrial settings (where the Mean Absolute Error typically needs to be controlled within 5% and the Root Mean Square Error below 1%). Nevertheless, it demonstrates distinct advantages over other comparative models, highlighting its effectiveness and superiority.</p>