Real-time traffic turning flow forecasting under few-shot and zero-shot scenarios by decoder-only model

Traffic Turning Movement Flows (TMFs) forecasting is crucial for optimizing traffic management and mitigating congestion in urban areas. However, traditional forecasting models struggle in zero-shot and few-shot scenarios where limited or no historical data is available. This research evaluates the effectiveness of decoder-only models, specifically TimesFM, for real-time TMFs forecasting and compares them with encoder-decoder models, such as Temporal Fusion Transformer (TFT), DeepAR and Long short-term memory (LSTM). We enhance TimesFM by integrating Empirical Mode Decomposition (EMD) for feature extraction and Random Forest (RF) for incorporating external covariates, addressing its interpretability limitations. Experimental results demonstrate that TimesFM, particularly when augmented with EMD and RF, outperforms traditional encoder-decoder models in zero-shot forecasting while maintaining competitive accuracy in few-shot scenarios. Additionally, we assess model performance across different weather conditions and forecast horizons, using Dynamic Time Warping (DTW) to quantify temporal misalignment. Our findings indicate that a hybrid approach that combines EMD and RF improves forecasting accuracy, especially under varying weather conditions and longer forecast horizons, making them better suited for deployment in dynamic environments.