Research On Interpretable Short-term Traffic Flow Forecasting Model Ing Method ——The Case Of Guiyang City

As the core part of Intelligent Transportation System（ITS）,short-term traffic flow prediction has received extensive attention with the rapid development and deployment of ITS.As the increasing popularity of traffic road sensor application,the types and quantities of traffic data have been greatly enriched,and in the meantime data-driven forecasting methods have gradually become the mainstream.As a representative result of data-driven methods,deep learning models have been widely used in traffic flow prediction because they can well capture spatial correlation.However,this kind of model is usually a black box with poor interpretability and fails to explain the causality between historical traffic observation data and prediction results.But in practice,the interpretability of traffic flow prediction results has important application value,which can enable managers to understand the causes of traffic jams,quantify the impact of each road section on traffic conditions,and predict the dynamic changes of traffic conditions over time,so that thy can make more reasonable decisions.Aiming at the interpretability of short-term traffic flow prediction results,this paper proposed an interpretable modeling method and conducted experiments on the Guiyang City Traffic Data Set as an example.This method divided the traffic time series into two parts: trend and fluctuation.First,the trend was about to be extracted,and then after the trend was eliminated,the residual series was tested for stationarity and randomness.After passing the tests,the autoregressive model was used to fit the residual series to determine how many lags in the history are related to the current traffic value of each intersection;Then the data was organized in a specific structure and fed into the GMDH network.The final experimental results were obtained after the prediction results by GMDH plus the eliminated trend.As a self-organizing network,GMDH don’t need to predefine input nodes,and the causality between input and output can be expressed as a polynomial function with finite elements,which greatly improves the interpretability of the whole model.To verify the advance in performance and interpretability of the proposed method,this paper used 3 experimental indicators to check the results created by completely different reference function types and compared the final best results with the LSTM baseline method and ARIMA+GARCH model.And the conclusion explained how the interpretability of the model was improved.The results show that the best reference function types to express the correlation between input and output are linear and cubic polynomial.The method proposed in this paper is better than LSTM and ARIMA+GARCH model in terms of accuracy and robustness,and the interpretability is also improved.