Supply And Demand Forecast Of Online Driver And Passenger Billing Based On Time Series

With the rise of the mobile Internet,the network car came into being.In the business scenario of the network car,the number of online drivers in different areas of a particular city varies greatly,and the amount of passengers issued is also very different.In order to solve the problem of imbalance between supply and demand between regions,it is first necessary to have an accurate prediction of the supply-demand ratio of each region.The supply-demand ratio of a particular area changes with time,is related to many external factors,has periodicity,and causes abnormal values in the sequence due to extreme events such as holidays and bad weather.When using the traditional third-order exponential smoothing,the abnormal points are not well predicted.Therefore,it is of great significance to take into account the amount of bills issued and the amount of orders received.Traditional machine learning models and cyclic neural networks have achieved certain results in the study of time series prediction.Based on these two directions,this paper designs a combined forecasting model based on the specific business background,and it has greatly improved compared with the existing methods.First of all,we obtain the supply and demand ratio of multiple regions in multiple cities at different time points as well as historical information such as regional location and weather.The primary features such as outlier processing,normalization,one-hot,and continuous feature discretization are performed on the corresponding features.Then,the processed features are further processed by two schemes.In the first scheme,we input the supply-demand ratio sequence into the holt-winter model of the third-order exponential smoothing and the first-order exponential smoothing,respectively,to solve the model parameters.The results of the linear weighting of the two models are then taken as a final feature.In the second scheme,we first integrate the data including regional features,weather features,solar features,window sequence features,etc.into the form of [batch,sequence,feature],and integrate the integrated data as input and output to train one layer.A self-encoder consisting of LSTM.The LSTM includes an input gate,an output gate,and a forgotten gate.The best model is obtained by iterating forward and backward propagation algorithm to parameter convergence.The encoder portion of the model is then encoded to encode the sequence features and the output of the encoder is taken as the final plurality of features.The data output from the previous two schemes is concat into a complete vector,combined with other features,input into the xgboost model,and iteratively obtains the final model parameters.Finally,the smape indicator is used to measure the forecasting effect of supply and demand on the test set.All data about supply and demand for several months before and after the acquisition,input to the above model,and train the model to parameter convergence.The experimental results show that the first-order exponential smoothing and the third-order exponential smoothing are weighted average,and the self-encoder is combined with xgboost for prediction.Significantly improved than the traditional single time series prediction method.