<p>Multivariate time series forecasting aims to model dynamic dependencies among temporal variables for accurate future predictions. Despite advances in deep learning, two key issues persist: ineffective dynamic feature weighting leading to overfitting and poor generalization and the inability of LSTMs to integrate multi-scale temporal features, limiting their capacity to capture both short- and long-term patterns. To address these challenges, we propose the Hadamard product LSTM (HPLSTM) model, which integrates an adaptive hierarchical feature fusion mechanism. Specifically, shallow and deep LSTM layers are leveraged to extract low-level and high-level temporal features, respectively—enabling the model to capture both short-term variations and long-term dependencies inherent in the time series. To further enhance the model’s representational capacity and adaptively emphasize informative patterns, Hadamard products are employed to conduct fine-grained feature selection and dynamically modulate the importance of hidden states across different network layers. This dedicated mechanism empowers the model to highlight crucial temporal signals while suppressing irrelevant or redundant information, thereby yielding substantial improvements in overall forecasting performance. Experimental evaluations on nine real-world datasets demonstrate that HPLSTM consistently outperforms baseline models. Compared with state-of-the-art approaches in the field, our proposed HPLSTM achieves an average reduction of 8.9% in Mean Squared Error (MSE) and 3.9% in Mean Absolute Error (MAE), validating its superiority in multivariate time series forecasting tasks.</p>