<p>Due to the complexity of reservoir geological conditions and the limited availability of geological data, the history matching process often encounters the issue of multiple possible solutions. Relying solely on a single geological model for history matching cannot ensure an accurate representation of the subsurface reality. To address these challenges, this paper proposes a machine learning-based method for automated history matching using multiple models. By employing deep autoencoders to reduce the dimensionality of multiple random geological models, and applying the WSO-KMeans clustering method to analyze the dimensionality-reduced results, we generate several initial model sets. These models are then used with the SPSA optimization algorithm for history matching and future dynamic forecasting, providing a range of predictions for future development. Practical applications demonstrate that the reconstructed results obtained through deep autoencoder-based dimensionality reduction can retain most of the key features of the original models. The WSO-KMeans clustering method effectively selects initial models that are representative of these features, facilitating accurate history matching and multi-model prediction. The resulting predictions are no longer confined to a single dynamic curve, but instead consist of a series of curves representing various development scenarios. The method provides a more in-depth study for multi-initial model selection and automatic history matching, with lower computational costs compared to traditional history matching methods, e.g., the time used for history matching using models with reduced dimensions is 70% lower than in full dimensions, enabling numerical simulations to provide a more reliable and convenient predictive reference for formulating reservoir development plans, especially for reservoirs that conform to Gaussian distributions.</p>