关键词: 工程物资配送, 路径优化, 行程时间, 端到端预测后优化, 双向长短时记忆网络

Abstract: This paper studies the optimization problem of distribution route for major engineering material based on the end-to-end predict-then-optimize (EPO) approach. First, a route optimization model for engineering material distribution is developed with the objective of minimizing total travel time. Then, considering that travel times on construction routes are highly affected by factors such as weather and geological conditions, a travel-time prediction model based on bidirectional long short-term memory (BiLSTM) neural network is proposed, along with a multi-source temporal data fusion mechanism. Subsequently, a differentiable training framework based on an upper bound of decision loss is developed to address gradient backpropagation stability by relaxing the non-convex loss function. Finally, a real-world case study from a railway construction area is used to validate the proposed the EPO-BiLSTM model. The results demonstrate that the EPO-BiLSTM model significantly outperforms benchmark models in terms of travel time reduction, routing planning stability, and robustness.

Key words: engineering material distribution, route optimization, travel time, end-to-end predict-then-optimize, bidirectional long short-term memory network (BiLSTM)