This research aims to address critical challenges in multivariate time series forecasting: capturing complex temporal patterns and overcoming the exponential decay issue in the long-term memory of Long Short-Term Memory (LSTM) models. To address these, we propose a new deep learning model called EL-LSTM. The model combines a simplified Leaky Integrate-and-Fire (LIF) neuron model with LSTM and significantly enhances the model’s ability to capture complex temporal dependencies through local and global point attention mechanisms. The EL-LSTM model can effectively retain key information. We have validated the effectiveness of the model in multivariate time series prediction tasks, especially in the fields of finance and traffic flow forecasting. Experimental results show that the EL-LSTM model has achieved significant improvements in forecasting accuracy, particularly when dealing with data in concept-drift environments. This study offers a new perspective for the field of time series prediction and demonstrates the potential for handling complex time series data in practical applications.