Research On Elevator Group Dispatching Base On Deep Reinforcement Learning

Modern buildings are widely equipped with multiple elevators.The elevator dispatch system should adopt passenger-flow-aware group scheduling strategy to coordinate these evevators,so as to reduce not only the passengers' time on waiting for and taking elevator,but also the energy consumptionMost established elevator group scheduling algorithms adopt a single optimization objective,and therefore lack for flexibility in concern of the change of passenger flow.This paper proposes a more systematic and flexible solution,resulting in an elevator scheduling algorithm which takes into account a variety of optimization objectives and exploits the potential of artificial intelligence.The proposed solution is mainly divided into three modules:a passenger flow pattern recognition module,a passenger flow data set construction module for various traffic patterns,and a training module for evlevator group scheduling which is based on deep reinforcement learning With the help of a sliding window on the passenger flow data and K-means clustering algorithm,the pattern recognition module identifies different patterns from the passenger flow.Given a specific passenger flow pattern,the dataset construction module generates simulated passenger flow data,which is then used to train the elevator group scheduler.The core of the training module is E-TD3,a revision of the famous deep reinforcement learning algorithm TD3..Multiple group scheduling policies are therefore respectively tuned for different passenger flow patterns Preliminary experiments show that:1)E-TD3 significantly outperforms the baseline DQN algorithm in that it yields better performance under all passenger flow patterns,and it takes far less time in training;2)E-TD3 has achieved good results in Baidu elevator dispatch competition,which shows that the algorithm has certain competitiveness.