Logistics demand prediction based on gated recurrent unit-cross attention mechanism

Purpose This paper aims to improve the accuracy of logistics demand forecasts, especially taking into account the impact of previous time steps on current logistics demand, to provide a basis for the government to integrate logistics resources, plan infrastructure development and formulate logistics industry policies. Design/methodology/approach In this paper, a method combining gated recurrent unit (GRU), cross-attention mechanism and sliding window method is proposed for logistics demand forecasting. With this method, feature information and time series information can be extracted from the hidden states of the GRU and the combination of parameters can be optimized using the grid search method. Findings The experimental results show that the prediction performance of this method outperforms that of Transformer, GRU- feature attention, GRU-temporal attention, long short-term memory (LSTM) and LSTM-Attention, as evidenced by lower mean square error (0.006), root mean square error (0.075) and smaller mean square absolute error (7.768%). Research limitations/implications Although the method performs well in experiments, it still has some limitations. For example, the complexity of the model is high and the computational cost is large and, the effect of the model may depend on specific data sets and parameter settings and the generalization ability needs to be further verified. In addition, the robustness of the model to anomalous data needs to be further investigated. Practical implications This study significantly improves the accuracy of logistics demand forecasting by introducing the cross-attention mechanism and optimizing the parameter combinations. This not only provides a scientific basis for the government to formulate relevant policies but also provides new ideas and technical means for time series forecasting in other fields. Originality/value In this paper, the authors innovatively apply the cross-attention mechanism to both feature variables and time series and optimize the model parameters through the sliding window method and the grid search method so as to improve the accuracy of logistics demand forecasting.