自动化立体仓库退库货位优化问题及其求解算法OA北大核心CSTPCD

The efficient retrieval handling of storage products in automated storage and retrieval system ( AS/RS ) is the key for just-in-time production in industrial production systems. In a production-oriented AS/RS, the amount of a specific product to be retrieved is often less than a pallet load. On completion of each retrieval request, the remaining products on the pallet should be relocated to a specific rack location. In current practice, after retrieving certain items, a non-empty pallet is typically returned to the originally designated location. It is of significant importance to dynamically re-assigning the storage locations of non-empty pallets after each retrieval in order to improve retrieval operational efficiency as well as energy consumption. In this study, a retrieval-oriented storage relocation problem is investigated to minimize energy consumption. Based on the characteristics of the acceleration and deceleration motion of the stacker crane and the crane's dual operation mode, an energy consumption model of AS/RS during retrieval processes is first established. Then the relocation of non-empty pallet after each retrieval request is taken as an operational decision and the constraints including rack stability and maximum storage capacity of each rack location are considered, and an optimization model of retrieval-oriented storage relocation is built with the objective of minimizing the total energy consumption during the retrieval processes. In consideration of the NP-hard characteristic of the optimization model, this paper proposes an optimization framework based on deep reinforcement learning algorithm. Within this framework, the multi-dimensional state is designed based on the real-time storage information as well as the retrieval information and the action is constructed by the retrieval-oriented storage relocation. A Markov decision process model for retrieval-oriented storage relocation is then established and the dueling double deep Q-networks ( D3QN ) algorithm is employed to obtain the optimal storage relocation of non-empty pallet after each retrieval request. In D3QN algorithm, the multi-dimensional state of AS/RS is put into the dueling network to predict main Q value as well as target Q value and the network training mechanism in the double deep Q network algorithm is introduced to train both main network and target network. When the network model converges to the optimal value function, the intelligent agent finds an optimal rack location for the non-empty pallet after each retrieval request based on the real-time state of AS/RS, so as to minimize energy consumption demanded by the stacker crane. Finally, the experimental cases of different scales are designed under the variability of rack shapes and product retrieval frequency. A set of sensitive analysis experiments are conducted to determine the hyper-parameters of D3QN, including training learning rate, the size of training batch, the number of neurons in the hidden layer and network update frequency. To evaluate the effectiveness of the proposed algorithm, extensive comparisons between D3QN, DDQN and DQN are made under different cases. Our experimental results show D3QN algorithm obtains better solution performance in dealing with the large-scale retrieval-related relocation cases. Moreover, a comparative analysis of the proposed relocation method and the method of returning to the original storage location is conducted under different cases. Our experimental results show the proposed relocation method significantly reduces total energy consumption during retrieval processes. Our study considers the sequence of retrieval requests as given rather than decision variables. If retrieval sequencing is incorporated with retrieval-oriented storage relocation, further improvement on retrieval efficiency and energy consumption may be achieved. Thus, further investigations are needed on the integrated optimization of retrieval-oriented storage relocation and retrieval sequencing for AS/RS.