MMP-Informer: an informer-based traffic flow prediction model by leveraging the previous cycle's max-min peaks

Abstract

Accurate traffic flow prediction is the core foundation of Intelligent Transportation Systems (ITS) and urban traffic dynamic optimization. However, existing mainstream prediction models have three critical limitations: poor cross-road transferability, mandatory separate training for individual road sections, and insufficient mining of historical traffic peak information, which jointly restrict prediction accuracy and engineering scalability in practical urban traffic management. To address these challenges, this article proposes the Multi-road Multi-Peak-aware Informer (MMP-Informer), an improved Informer-based long-term time series prediction model with a novel dual-peak-aware attention mechanism. This core module adaptively locates historical maximum and minimum traffic peaks via an adaptive window matched to input sequence length, introduces learnable parameters to weight the influence of different peaks on future traffic flow, and realizes single-model end-to-end simultaneous multi-road traffic flow prediction. Meanwhile, the model retains the encoder-decoder architecture of the original Informer, integrates self-attention distillation in the encoder to reduce computational overhead, and adopts a generative decoder paradigm to avoid error accumulation in multi-step prediction. Experiments on the 2018 University of Technology Sydney (UTS) multi-source traffic dataset show that MMP-Informer consistently outperforms state-of-the-art baselines including SCINet, NsTransformer, Transformer, and original Informer across 24-, 48-, and 96-h prediction horizons. For 24-h short-term prediction, it achieves a Mean Absolute Error (MAE) of 25.43 and Mean Absolute Percentage Error (MAPE) of 11.55%; for 96-h long-term prediction, it reaches an MAE of 31.21 and MAPE of 12.93%, where the MAE is 2.51 lower and MAPE is 1.97 percentage points lower than the original Informer, respectively. Ablation tests verify that the dual-peak-aware design significantly enhances the model’s adaptability to both low-traffic and high-traffic scenarios, balancing prediction accuracy and long-term forecasting stability. Supplementary experiments on merged Electricity Transformer Temperature hourly (ETTh1/ETTh2) datasets further confirm its stable and superior performance in non-traffic time series scenarios, with strong cross-scenario generalization ability. The proposed model eliminates the high cost of single-road separate training, breaks through the cross-road transferability bottleneck of traditional models, maintains stable accuracy in both short-term and long-term prediction, and has high practical engineering value for large-scale deployment in urban intelligent traffic management.