The material delivery route prediction problem aims to forecast the future delivery routes of couriers given a set of tasks. Due to the high non-linearity and complexity of vast historical data, as well as factors related to individual courier preferences, this task poses significant challenges. Most existing methods use deep neural networks based on supervised learning to capture behavior patterns from historical data. However, they often struggle with the dynamic nature of the data and the diversity of individual preferences. This paper proposes a new deep reinforcement learning framework that integrates Variational Autoencoders (VAE) to handle large-scale data features and incorporates dynamic embedding features to accurately reflect personal preferences. The framework is trained using Proximal Policy Optimization (PPO) for optimized policy. Experimental validation with two publicly available real-world urban delivery datasets from Cainiao Network and two datasets generated from cities across multiple countries shows that the proposed framework significantly outperforms seven existing prediction methods.

The rise of multi-industry chain networks has transformed traditional industrial operations into complex systems with intricate cross-layer resource dependencies and company group structures. Despite enhancing operational efficiency, these networks are increasingly vulnerable to task surges—sudden spikes in resource demand that propagate across layers, and trigger cascading failures. Existing emergency response strategies, designed for single-layer networks, fail to address the multidimensional challenges of maintaining critical company stability and ensuring overall network robustness. To address this gap, this paper introduces the Multi-Modal Collaborative (MMCol) algorithm, which incorporates: (1) A dual-dimension collaboration framework enabling coordination across intra-layer vs. inter-layer and single company vs. company group, (2) adaptive inter-layer collaboration coefficients for dynamic resource allocation and stability preservation, and (3) a hierarchical decision strategy that reduces computational complexity by treating company groups as collaborative units and by pre-selecting layers based on resource compatibility. Extensive simulations across various network scales and experiments on real industry chain datasets demonstrate that MMCol consistently outperforms state-of-the-art algorithms in key metrics including load balance, collaboration cost, and key entity risk protection. The performance advantages are particularly significant in complex networks that mirror real industrial structures, thereby confirming MMCol’s effectiveness in practical multi-industry chain environments.

With the high degree of integration of production capacity in the industrial field, the original form of single, linear, and vertical cooperation between different industrial chains has been broken, and a multiplex networked industrial chain has been formed. Traditional time series forecasting methods are often prone to fall into the trap of computational volume caused by long historical information and the problem of dimensional explosion caused by the mixing of redundant information in the face of multiple networked industrial chain capacity forecasting with large data volume and high information dimensionality. In this paper, we first propose an information decoupling technique based on the principle of time series decomposition to provide more accurate cyclical forecasting results for capacity forecasting. Secondly, this paper introduces a multi-scale dynamic aggregation network technique. This technique dynamically aggregates and predicts variables at different time scales. The combination of these two approaches is adept at capturing a wider range of local and global trends, thereby greatly improving the accuracy and robustness of forecasting models. In this paper, experiments are conducted to compare with the current mainstream time series prediction algorithms. The results show that in multivariate long time series, the error of our algorithm is reduced by 27.8%.