Global soil moisture is a vital climatic variable, playing a critical role in agriculture, water resource management, and climate research. While recent advancements in deep learning have markedly enhanced soil moisture prediction, these models often struggle to identify and integrate relevant physical features, leading to inaccuracies. To address this gap, we propose the AEDLSTM-HBV model, which integrates features from the Hydrologiska Byr & aring;ns Vattenbalansavdelning (HBV) model within a deep learning framework. This model employs an Attention-Enhanced Encoder-Decoder Long Short-Term Memory (AEDLSTM) network to improve the representation of soil moisture by effectively leveraging the fusion of physical features and original inputs. Experimental results on the LandBench1.0 dataset indicate that the AEDLSTM-HBV model surpasses state-of-the-art models. Notably, our model demonstrates robust performance in predicting soil moisture in permafrost and desert regions, achieving an average R2 improvement of up to 20% over the baseline AEDLSTM model. This study highlights the significant potential of integrating physical process features to enhance the predictive capabilities of deep learning models for global soil moisture. By incorporating and improving the representation of these complex physical processes, we can achieve more accurate and reliable predictions.