Method and system for delay prediction for scheduled public transport using multi architectural deep learning

A system for delay prediction for scheduled public transport, the system comprises:

• a user interface;
• one or more hardware processors;
• a memory in communication with the one or more hardware processors and the user interface, wherein the one or more first hardware processors are configured to execute programmed instructions stored in the one or more first memories, to:
• receive historical operation data corresponding to a plurality of vehicles deployed on a network of the scheduled public transport, wherein a route in the network having a queried vehicle for which an expected arrival delay is to be predicted on a current stop;
• transform the received historical operation data into temporal data and spatio-temporal data;
• obtain a spatial data from geographical information corresponding to the plurality of vehicles deployed in the network of the scheduled public transport;
• identify a first window of a set of stops from amongst the plurality of stops, wherein the set of stops are stops from where a trip of the queried vehicle has already completed;
• identify a second window of a set of vehicles of the current stop, wherein the set of vehicles of the current stop has passed through all the stops in the first window before the queried vehicle, wherein the set of vehicles are of a same mode that are in the second window based on type of vehicles, wherein the type of vehicles comprises:
• route-based vehicles, where the set of vehicles of the same mode has covered the same set of stops as the queried vehicle before the queried vehicle, and
• trip-based vehicles, where the set of vehicles of the same mode has covered the same set of stops in the first window considered as the queried vehicle before the queried vehicle;
• extract a plurality of spatial features for each stop in the first window, wherein the plurality of spatial features comprises a distance between two consecutive stops, and a station vehicle capacity;
• form a spatial feature vector of a dimension of twice the number of stops in the first window using the plurality of spatial features;
• extract a plurality of temporal features for each stop in the first window covered by vehicles in the second window, wherein the plurality of temporal features comprises a scheduled travel time between all consecutive stops in the first window before the current stop, a scheduled dwell time at a stop, a scheduled interval with preceding vehicle, a scheduled number of stops between a two consecutive stops in the first window, an actual travel time, an actual dwell time, and an actual interval with preceding vehicle for all consecutive stops in the first window;
• form a temporal feature vector using the plurality of temporal features, wherein the vehicle is chosen from the second window to form the temporal feature vector and wherein each of the temporal features are computed for the current vehicle, the route-based and/or trip-based vehicle, for each stop in the first window;
• form a spatio-temporal feature vector using a plurality of spatio-temporal features, wherein the plurality of spatio-temporal features is stacked into cubes, wherein the plurality of spatio-temporal features comprises an arrival delay and a departure delay for each of the plurality of vehicles and for each of the plurality of stops;
• train a first architecture of Fully Connected Neural Network (FCNN), using the spatial feature vector;
• train a second architecture of a Long Short Term Memory (LSTM) network using the temporal feature vector;
• train a third architecture of a 3D Convolutional Neural Network using the spatio-temporal feature vector, wherein a function of the spatial feature vector, the temporal feature vector and the spatio-temporal feature vector represented as a standard regression problem given by f(xS, xT, xST)=y;
• training a regressor network with respect to information flow between first, second and third architecture networks to obtain a loss function wherein loss is reduced using target variable “y” and the predicted variable “y” obtained from the standard regression problem using
• and constructing a feature vector for plurality of future inputs and feeding the regressor network with the feature vector for plurality of future inputs;
• train a plurality of final encoding models provided in a plurality of combinations of the first architecture, the second architecture and the third architecture, wherein encodings from the first, second and third architecture networks are combined together, wherein the plurality of combinations comprises seven combinations including six combinations have permutation dealing with connecting the first architecture, the second architecture and the third architecture in series and one combination has two models in parallel and one in series;
• choose a final encoding model from amongst the plurality of final encoding models based on accuracy of each of the plurality of final encoding models to predict the delay for the queried vehicle at the current stop, wherein a mean accuracy error is calculated to check the accuracy of the plurality of final encoding models and wherein the loss function obtained from the regression network is used to evaluate the plurality of final encoding models;
• and
• predicting a delay for the queried vehicle at the current stop using the final encoding model.

One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:

• receiving, via a user interface, historical operation data corresponding to a plurality of vehicles deployed in a network of the scheduled public transport, wherein a route in the network having a queried vehicle for which an expected arrival delay is to be predicted on a current stop;
• transforming the received historical operation data into temporal data and spatio-temporal data;
• identifying a first window of a set of stops from amongst the plurality of stops, wherein the set of stops are stops from where a trip of the queried vehicle has already completed;
• obtaining a spatial data from geographical information corresponding to the plurality of vehicles deployed in the network of the scheduled public transport;
• identifying a second window of a set of vehicles of the current stop, wherein the set of vehicles of the current stop has passed through all the stops in the first window before the queried vehicle, wherein the set of vehicles are of a same mode that are in the second window based on type of vehicles, wherein the type of vehicles comprise:
• route-based vehicles, where the set of vehicles of the same mode has covered the same set of stops as the queried vehicle before the queried vehicle, and
• trip-based vehicles, where the set of vehicles of the same mode has covered the same set of stops in the first window considered as the queried vehicle before the queried vehicle;
• extracting a plurality of spatial features for each stop in the first window, wherein the plurality of spatial features comprises a distance between two consecutive stops, and a station vehicle capacity;
• forming a spatial feature vector of a dimension of twice the number of stops in the first window using the plurality of spatial features;
• extracting a plurality of temporal features for each stop in the first window covered by vehicles in the second window, wherein the plurality of temporal features comprises a scheduled travel time between all consecutive stops in the first window before the current stop, a scheduled dwell time at a stop, a scheduled interval with preceding vehicle, a scheduled number of stops between a two consecutive stops in the first window, an actual travel time, an actual dwell time, and an actual interval with preceding vehicle for all consecutive stops in the first window;
• forming a temporal feature vector using the plurality of temporal features, wherein the vehicle is chosen from the second window to form the temporal feature vector and wherein each of the temporal features are computed for the current vehicle, the route-based and/or trip-based vehicle, for each stop in the first window;
• forming a spatio-temporal feature vector using a plurality of spatio-temporal features, wherein the plurality of spatio-temporal features is stacked into cubes, wherein the plurality of spatio-temporal features comprises an arrival delay and a departure delay for each of the plurality of vehicles and for each of the plurality of stops;
• training a first architecture of Fully Connected Neural Network (FCNN), using the spatial feature vector;
• training a second architecture of a Long Short Term Memory (LSTM) network using the temporal feature vector;
• training a third architecture of a 3D Convolutional Neural Network using the spatio-temporal feature vector, wherein a function of the spatial feature vector, the temporal feature vector and the spatio-temporal feature vector represented as a standard regression problem given by f(xS, xT, xST)=y;
• training a regressor network with respect to information flow between first, second and third architecture networks to obtain a loss function wherein loss is reduced using target variable “y” and the predicted variable “y” obtained from the standard regression problem using
• and constructing a feature vector for plurality of future inputs and feeding the regressor network with the feature vector for plurality of future inputs
• training a plurality of final encoding models provided in a plurality of combinations of the first architecture, the second architecture and the third architecture, wherein encodings from the first, second and third architecture networks are combined together, wherein the plurality of combinations comprises seven combinations including six combinations have permutation dealing with connecting the first architecture, the second architecture and the third architecture in series and one combination has two models in parallel and one in series;
• choosing a final encoding model from amongst the plurality of final encoding models based on accuracy of each of the plurality of final encoding models to predict the delay for the queried vehicle at the current stop, wherein a mean accuracy error is calculated to check the accuracy of the plurality of final encoding models and wherein the loss function obtained from the regression network is used to evaluate the plurality of final encoding models;
• and
• predicting a delay for the queried vehicle at the current stop using the final encoding model.