AutoVLCNN: Automated Variable-Length Convolutional Neural Networks for Traffic Flow Prediction

Traffic flow prediction is a fundamental component of intelligent transportation systems (ITS), playing a key role in optimizing management, reducing congestion, and improving efficiency. Deep learning (DL) has been successfully and widely applied to various traffic flow prediction scenarios in recent years, but existing models heavily rely on manual design of neural architectures (NAs) and hyperparameters (HPs), and lacking generalization capability. To address these issues, this paper proposes an automated variable-length convolutional neural network (CNN) design method for traffic flow prediction called AutoVLCNN, which jointly optimizes NAs and HPs to obtain the optimized model. In AutoVLCNN, an efficient hybrid variable-length encoding strategy is designed from a block perspective to characterize the topological structures and structural parameters of four different blocks in CNN, which are encoded using integer and continuous variables in the genetic algorithm (GA) framework. Furthermore, novel crossover and mutation operations are carefully designed to evolve both NAs and HPs, updating variable-length individuals. The experimental results demonstrate that AutoVLCNN achieves remarkable performance on the PeMS10,WebTRIS, and JxSTKL-TF datasets, with RMSE of 32.5572, 37.2595, and 12.7892, R2of 0.9370, 0.9654, and 0.8219, MAE of 22.8690, 25.4670, and 8.3796, and MAPE of 12.1039, 21.2441, and 64.5051, respectively. Its performance is notably superior to eight state-of-the-art manually designed methods and two automated DL-based methods.