A METHOD FOR IMPROVING A FLOW OF RAILWAY VEHICLES IN A RAILWAY STATION

Technical field

This invention relates to a method and a system for improving a flow of railway vehicles in a railway station including a plurality of tracks.

Background art

In the sector of planning a railway station including various tracks it is necessary to plan the station by associating with each incoming railway vehicle a corresponding arrival track and a corresponding departure track. Moreover, in some situations, it may also be necessary to take into consideration the distance between the associated tracks, in order to limit the extent of the movement of the passengers for changing railway vehicles. For this reason, planning must be carried out in the best way. Initially, this planning was carried out manually by the station manager, who, starting from initial data established an initial planning. This solution was obviously not very optimised, due to the reduced capacity of the person in charge of weighing up optimum situations in conditions as complex as railway stations with many tracks.

Moreover, such solutions were not resilient to unforeseen changes to the station that modified the condition of the station, making compliance with the original planning impossible.

In order to overcome the drawbacks of the manual management of the station, methods have been implemented for improving the flow of railway vehicles. For example, patent documents CN110341763A, GB2508508A and CN104875774A describe solutions for the dynamic planning of high speed trains. Patent CN110341763A describes a smart optimizer, which picks up the initial parameters of a certain section of a train from a static database, and the dynamic parameters relative to the operation of the train, collected in real time through a data collector. Based on this data, the optimiser re-organises the flow of the train following any unforeseen events detected on the dynamic parameters.

Patent GB2508508A describes a solution of a timetable re-planning apparatus for managing operation of the vehicles (for example, a train after a delay) and performing the timetable re-planning according to the circumstances. The apparatus comprises information management means for determining the level or severity of a conflict affecting the traffic of the vehicles on the basis of information regarding the conflict which has occurred at the site. The management of the priority level consists in generating information on the priority level of the vehicle for each vehicle in the re-planning of the timetable concerning the conflict, based on realtime information on the passengers, to define at least one rate of congestion for each vehicle or a rate of congestion of the passenger station in each passenger station.

On the other hand, patent CN104875774A describes a method for adjusting the delay based on an operating diagram of the urban railway transit. The method comprises the acquisition of operating parameters of the train, wherein the operating parameters of the train comprise information such as the data of the train operating diagram, the minimum stationary time at the station, the operating time of the minimum interval and the like. The method comprises establishing a module for adjusting the delay of the operating diagram based on the distribution of the arrival and departure time of the train using the operating parameters. The method comprises the calculation of the model for adjusting the delay of the operating diagram of the train on the basis of odds ratio formulae.

These methods, although they solve the problems linked to a manual calculation of the planning and its adaptation, are imprecise because they are not very flexible to unforeseen events which can occur and are therefore not very reliable.

Other methods present in the prior art, which however have the same drawbacks, are, for example, described in patent document CN1 11376954B.

Aim of the invention

The aim of this invention is to provide a system and a method for improving a flow of railway vehicles which overcome the above-mentioned drawbacks of the prior art.

Brief description of the drawings

These and other features of the invention will become more apparent from the following description of a preferred, non-limiting example embodiment of it illustrated in the accompanying drawing which schematically illustrates a system for improving a flow of railway vehicles in a railway station.

Detailed description of preferred embodiments of the invention

Said aim is fully achieved by the method and system according to the invention as characterised in the appended claims.

According to an aspect of the invention, the invention provides a method for improving a flow of railway vehicles in a railway station including a plurality of tracks. It should be noted that the method and the system according to the invention could also be applied for improving the flow of other vehicles other than trains, such as, for example, automobiles, lorries, ships and aircraft; however, this description will be focussed, for convenience, on the application which is currently of greatest interest, that is to say, that of railway vehicles, but without thereby limiting the scope of this patent document.

The method comprises a step of receiving stretch data, including, for each railway vehicle, a reference time of arrival in the station and a reference time of departure from the station. The data processing is therefore representative of the initial planning of the trains.

The method comprises a step of receiving status data, representing a status (that is, a configuration) of the station at a given instant in time. The status data may also include commercial restrictions which are indicated by the operator of the railway station, for example, but without limiting the scope of the invention, a certain priority on trains of certain types with respect to others of other types. The status data identifies, for example, but without limiting the scope of the invention, the number of tracks available and the position of the diverters at the entrance to the station.

The status data represents the occupation of the stabling tracks and the routes which can be travelled by said railway vehicles inside the railway station.

According to an embodiment, the method comprises a step for receiving railway network data. The railway network data include, for each railway vehicle, a corresponding planned arrival time at the railway station. It should be noted that said time of arrival at a station may coincide with the time planned by the stretch data or may differ from said time, for example in the case of delays.

According to an embodiment, the method comprises a step of supplying status data and railway network data to a machine-learned model. The machine-trained model is programmed to generate programming data. The programming data includes one or more of the following items of information, relating to one or more railway vehicles: a stabling track; an arrival and/or departure route, identifying an entrance and/or exit path with respect to the stabling track; an arrival or departure time.

It should be noted that the status data received in the receiving step are the same status data which is then used by the machine-learned model to generate the programming data.

The machine-learned model is trained, starting from the status data and the railway network data, to derive the programming data to reduce a deviation of the arrival or departure time derived with the programming data relative to the reference arrival or departure time.

The fact that the status data represent the occupation of the stabling tracks and that the programming data include, for one or more railway vehicles, a stabling track and an arrival and/or departure route, identifying an entrance and/or exit path with respect to the stabling track are features which collaborate to allow an efficient reprogramming, considerably improving the efficiency of the method. In effect, a method which does not have information on which stabling tracks are available is penalised, because it is not able to perform a reassignment of the tracks, not knowing which ones are available.

Moreover, once the stabling track has been reassigned, following processing with the trained machine-learned model, providing information regarding the path which the train must follow, that is to say, the entrance and exit, is central for assessing the deviation between initial programming and new programming, so assessing how efficient the reprogramming is. For this reason, the synergy of these features has the technical effect, that is to say, the advantage of preventing the occurrence of interference between paths which would delay reaching the new stabling track, thereby improving the effectiveness of the planning.

According to an embodiment, the method comprises a step for access to initial programming data.

The programming data represents a route of each train within the railway station. The programming data can include the following information:

- stabling track in station;

- a route ranging from an entrance protection signal of the station (which identifies the entrance into the station) to the stabling track;

- a route ranging from the stabling track to an exit protection signal of the station (which identifies the exit from the station) to the stabling track;

- direction of the train (also indicated in the stretch data).

The programming data are, in other words, the initial planning which associates to each railway vehicle a corresponding stabling track and the transit route in the railway station. This programming is performed previously, and may be performed by means of a processor or manually by an operator.

Preferably, in order to access the initial programming data, the method comprises an initialisation step, wherein, starting from the stretch data and from the status data at an initialisation instant, the processor derives the initial programming data, which comprise, for each railway vehicle, a stabling track, an entrance route, identifying an entrance path of the station to the stabling track, and/or an exit route, identifying a path from the stabling track to the exit of the station.

The method may comprise a step of receiving operating data, representing an external disturbance which varies the status data. The operating data are, in other words, data which identify a certain type of unplanned event, for example data representing the interruption of a track, data representing the fault of a train or other events of different types. These operating data can therefore determine a variation of the status data, and, therefore, consequently, a variation of the programming data derived from the machine-learned model.

The method comprises a learning step. The learning step comprises a step of generating simulation data, including simulated railway network data and simulated status data.

The method comprises a step of sending simulation data to the machine- learned model.

The method comprises a step of processing simulation data. The method comprises a step of identifying a first hypothesis of programming data.

The method comprises a step of calculating an assessment value for the hypothesis of calculated programming data. The assessment value represents an overall delay with respect to the reference departure times or movements of passengers inside the railway station.

The simulation step comprises a step of identifying a second hypothesis of programming data and calculating the corresponding assessment value.

According to an embodiment, the machine-learned model is trained by varying the simulation data, for being trained relative to a plurality of states of the railway station and/or to a plurality of railway network data of the railway vehicles in the railway station.

The presence of a learning step, wherein the method feeds with simulated examples allows a very high precision and reliability of the method to be obtained.

Preferably, the simulation module has access to has a historical data archive, including status data and railway network data associated with cases which have actually occurred.

According to an embodiment, the method comprises a step of generating simulation data, in a simulation module of the processor, which generates simulated status data and simulated railway network data which do not correspond to historical data but which are in any case simulated to further increase the precision of the machine-learned model.

This feature makes it possible to further increase the precision of the method by providing simulations of unforeseen events which have not yet been tested but which, potentially, can occur.

Preferably, the machine-learned model is trained through reinforcement learning techniques.

Preferably, the machine-learned model includes a trained neural network. This guarantees further precision and reliability in identifying the optimum solution.

According to an embodiment, the initialisation step described above comprises a step of access to a target function, determined starting from said stretch data and from said status data. It should be noted that the term "target function" means a function which links the programming data with a target parameter, which is selected by the user of the method on the basis of the relative business. For example, the target parameter could be a value of movement of passengers for changing railway vehicles or an overall delay value of the railway vehicles.

According to an embodiment, the target function is defined by a vector of the unknowns, which represents the programming data to be calculated, and by a matrix, derived on the basis of the status data and the stretch data. The product between the unknown vector and the matrix provides a passenger movement value for changing the railway vehicle or an overall delay value of the railway vehicles (that is to say, a value of the target parameter).

In the initialisation step there is a step of determining the programming data by minimising the target function, in order to improve the assessment value.

This allows starting with the programming data already optimised, on which to intervene only in the case of unforeseen events.

According to an aspect of the invention, the invention provides a further embodiment of a method for improving the flow of a plurality of railway vehicles in a railway station including a plurality of tracks. Said embodiment, like the one illustrated above, includes one or more of the following steps:

- accessing stretch data, including, for each railway vehicle, a corresponding reference time of arrival to the railway station and a reference time of departure from the station;

- receiving status data, representing a status of the station at an instant in time, said status data representing the following information:

- the occupation of the stabling tracks;

- the routes which can be travelled by said railway vehicles within the railway station;

- receiving railway network data (204), including, for each railway vehicle (101 ), a corresponding planned arrival time at the railway station;

- deriving a target function, which correlates the programming data with a target parameter; said target function being derived on the basis of the stretch data, railway network data and status data;

- minimising the target function in order to identify the programming data which determine the minimum value of the target parameter, that is to say, the minimum value of the overall delay or the minimum value of the movements of the passengers.

Also in this embodiment of the method there is an initialisation step, wherein the programming data are calculated with a target function which is determined on the basis of the stretch data and the status data, as described above.

Preferably, the target function is a quadratic target function. This makes it possible to obtain more reliable and responsive results.

According to an aspect of the invention, in both the embodiments of the method, the constraining data comprise one or more of the following items of information:

- the presence of a single railway vehicle on a track;

- specification on the length of each track;

- interval of time between occupation of a track by a railway vehicle and occupation of the track by a further railway vehicle which is greater than a minimum interval of time;

- type of each track between arriving and/or departing track.

According to an aspect of the invention, the invention provides a control system for improving the flow of a plurality of railway vehicles in a railway station including a plurality of tracks. The system comprises a processor, programmed to perform the steps of the method including any of the steps described with reference to the method for improving the flow of vehicles according to the invention.

In addition, the invention aims to also protect a computer program including instructions for performing the steps of the method including any of the steps described with reference to the method for improving the flow of vehicles according to the invention.

In order to clarify further, it should be noted that the method comprises two steps (that is to say, the method comprises steps for achieving two respective purposes), a step of static calculation of the programming data (also referred to as the initialisation step) and a dynamic calculation step, which calculates the programming data in a dynamic manner in response to receiving the railway network data.

According to a first embodiment, both the static calculation step and the dynamic calculation step are performed by means of the machine-learned model. According to a second embodiment, both the static calculation step and the dynamic calculation step are performed by minimising the target (quadratic) function.

On the other hand, according to a preferred embodiment, the static calculation step is performed by minimising the target (quadratic) function whilst the dynamic calculation step is performed by means of the machine- learned model.

The invention illustrates a method for improving a flow of railway vehicles 101 in a railway station 102 including a plurality of tracks 103.

The method is performed by a processor which receives a series of data aimed at enabling it to improve the flow of the vehicles. In particular, the method focuses on two main aspects of the flow improvement. The first aspect concerns the static improvement of the programming. “Static improvement” of the programming means the activity of optimising the programming performed initially, to identify, in the absence of unforeseen events, which is the optimum programming.

The second aspect concerns the dynamic improvement of the programming. “Dynamic improvement” of the programming means the activity of optimising the programming performed continuously, starting from initial programming data, for updating the programming in response to a variation of the stretch data or initial status data, for example a delay of a specific railway vehicle or a track which is not in use.

In order to perform these functions, according to the method the processor receives stretch data 201 , including reference arrival times in the station and reference departure times from the station for a plurality of railway vehicles. In other words, the stretch data identifies the stretches which must be travelled by the railway vehicles 101 and therefore define constraints relative to the time in which a specific railway vehicle must reach the station 102 and also when it must leave the station.

According to the method, the processor receives status data 202, representing a status of the station at an instant in time. The status data, according to an embodiment, are sent by a central server (called safe core) which detects all the conditions of the railway network and updates, preferably at a predetermined frequency, each processor, to allow it to optimise, according to the method described therein, the programming of the respective station.

More specifically, the status data represents the occupation of the stabling tracks and/or the routes which may be travelled along by said railway vehicles inside the railway station. Clearly, this information refers to a specific instant in time, in which the processor is calculating the optimum programming.

The status data are therefore significant since the processor knows which tracks and routes can be used and which, on the other hand, are currently occupied.

It should be noted that the status data 202 may include two types of information:

(i) information relating to the real-time occupation of tracks and routes, which is information which varies most frequently every update which is sent to the processor;

(ii) information more linked to structural constraints of the station which are invariable apart from unforeseen events which result in structural variations in the station (for example, interruption or closing of a track due to damage).

Thus, the status data 202 comprises constraining data, for example representing a layout of the railway station. The constraining data are data which physically represent the railway station and which include one or more of the following data (information): - number of tracks 103;

- type of railway vehicle 101 suitable for each track 103;

- maximum number of railway vehicles 101 present simultaneously on each track 103;

- maximum time interval for the connection intervals between railway vehicles 101 ;

- maximum distance which can be travelled by the passengers on the tracks 103;

- length of each track 103 (to assess whether a specific vehicle can enter the track);

- length of each railway vehicle 101 ;

- minimum interval of time between occupation of a track 103 by a railway vehicle and occupation of the track 103 by a further railway vehicle 101 ;

- type of each track 103 between arriving and/or departing track;

- complete route of each railway vehicle, within the railway station 102, from an entrance track to a departure track;

- list of crossings, including, for each crossing stretch the corresponding two routes which cross.

According to the method, the processor has access (receives, calculations, derives) to initial programming data 203. The initial programming data 203 includes, for each railway vehicle 101 , a corresponding predetermined stabling track 103 between said plurality of tracks 103 of the station 102 and/or an entrance route to the stabling track and/or an exit route from the stabling track. The initial programming data

203, as stated above, may be received by the processor or may be derived from the processor by means of the static improvement step of the programming, as described in more detail below.

According to the method the processor receives the railway network data

204, including, for each railway vehicle 101 , a corresponding planned arrival time at the railway station. The railway network data 204 are data which represent real railway network data in real time of the railway vehicles 101 . The railway network data 204 may represent a delay of one or more railway vehicles, when the planned arrival time at the railway station differs from the reference arrival time.

Preferably, the method is configured for updating the programming data with respect to the initial programming data if the railway network data 204 in effect represent a delay of at least one train, to avoid a waste of computational calculations in the absence of a variation of the initial conditions. In other words, according to an embodiment, the machine- learned model MAM is launched if, for at least one train, the planned arrival time at the railway station differs from the reference arrival time. However, this is not necessarily essential, as the machine-learned model MAM could still be launched at regular intervals.

The railway network data 204 are preferably received in real time. The railway network data 204 are communicated in real time by means of a remote connection with the processor. According to an embodiment, the railway network data 204 are received automatically in the processor, by means of an integration between systems for monitoring circulation (which may be on each railway vehicle 101 or outside the vehicles and communicating with the latter) and the processor whilst, according to other embodiments, they may be inserted manually by an operator who supervises the processor.

According to the method, the processor has access to a data archive 300. The data archive 300 comprises railway network data 204 and status data 202 which represent historical cases which have actually occurred, that is to say, delays of railway vehicles which have actually been tested or of statuses of the station which have actually occurred.

The data archive 300 is a container in which a plurality of cases are stored.

The method comprises a step 203' of calculating optimum programming data by means of a machine-learned model MAM. The machine-learned model MAM is trained to calculate the programming data 203’ which minimise a deviation of the arrival or departure time relative to the reference arrival or departure time.

It should be noted that, preferably, this system is conceived to minimise the effects that a deviation between the reference arrival time and the actual arrival time has on the delay which can be accumulated in the station. Preferably, the machine-learned model MAM is trained to calculate the programming data 203’ which minimise a deviation of the departure time relative to the reference departure time, due to the internal organisation of the station (that is, stabling tracks and routes). It should be noted that, in minimising the deviation between the departure time with respect to the reference time, the machine-learned model preferably takes into consideration a minimum stabling time for which the railway vehicle must remain on the stabling track, to allow the operations for loading and unloading the passengers. For this reason, if any reduction in the stabling time is evaluated this always done remaining above the minimum stabling time.

Moreover, according to other embodiments, it is possible that the machine-learned model MAM is trained to calculate the programming data 203’ which minimise a deviation between the movements of the passengers in the station with respect to the movement of the passengers in the station calculated with the initial programming data 203.

Advantageously, the method comprises a learning step. In the learning step, the processor receives simulation data 206, including simulated railway network data 204’ and simulated status data 202’.

The learning step comprises, for each value of the simulated railway network data 204' and the simulated status data 202', determining one or more hypotheses of programming data which allow the machine-learned model MAM to be trained. More in detail, the processor determines a first hypothesis of programming data. The processor calculates, for said first hypothesis of programming data, a value of the deviation of the arrival or departure time with respect to the reference arrival or departure time, for each of the railway vehicles. Said value defines an assessment value, which is used by the processor to determine the quality of the solution identified. Subsequently, the processor determines a second hypothesis of programming data. The processor calculates, for said second hypothesis of programming data, a value of the deviation of the arrival or departure time with respect to the reference arrival or departure time, for each of the railway vehicles. The processor, comparing the assessment value of the first hypothesis with that of the second hypothesis, learns to discriminate the solutions. In the same way, repeating this process with a plurality of hypotheses of programming data, of which it calculates, for each, the relative assessment value, the machine-learned model MAM increases its learning and therefore refines the capacity to find solutions, even in untried contexts, which reduce the assessment value.

This technique is most commonly known as reinforcement learning.

Advantageously, the method comprises a step of generating simulation data 206, in a simulation module of the processor. In other words, the simulation module is able to envisage new cases corresponding to new simulated railway network data 204’ and simulated status data 202’ (not necessarily tested and included in the remote archive 300), to allow the data archive 300 to be made even more flexible, sensitive and reliable.

According to an embodiment purely by way of example, the selection of the updated programming data in response to the predetermined external disturbance is performed by means of a trained neural network.

As stated above, the method may comprise a step of static improvement of the programming, wherein the processor has access to a target function, which is determined starting from said processing data and from said status data. Moreover, the target function is also determined on the basis of the target parameter which is to be minimised, that is to say, the overall delay of the trains or the movement of the passengers inside the station. This target parameter depends, in short, on a business choice, so the target function can vary with the variation of the needs expressed by the operator who uses the method.

According to an embodiment of the method, it is possible to indicate the target parameter, for example to allow a user of the method, in a critical manner, to assess whether it is more important to reorganise the station to minimise the delay or to minimise the flow of persons. This may be important when, over time, a transport service provider wants to vary the parameter with which to optimise the programming data. From a practical point of view, the method comprises an updating of the target function on the basis of the selection of the user, in such a way as to be able to calculate the new programming data in order to minimise the new target parameter. This aspect, although not an essential feature of the method, allows it to be flexible and versatile.

As well as being used for the initialisation step, the target function can also be used for the dynamic improvement step of programming. The target function may, however, be aimed at minimising a different target parameter, if it is used for the static optimisation step or for the dynamic optimisation step. For example, but without limiting the scope of the invention, in the static improvement step the target parameter is the interference between trains whilst in the dynamic improvement step the target parameter could be the total delay of the trains.

Advantageously, the target function is a quadratic target function. In other words, the solution of the target function is a solution of a problem defined by the term QUBO (Quadratic Unconstrained Binary Optimization). It will be understood that this is only an example and that other types of quadratic functions can also be used.

Moreover, it is further clarified that the approach which uses a quadratic target function is a different approach (alternative or combinable) to the approach which uses a machine-learned model (preferably with Reinforcement Learning techniques).

For example, a quadratic function is a function which comprises a binary vector which encodes the solution and a matrix which encodes, on the other hand, the problem to be resolved. Thus, more specifically, the matrix is determined on the basis of the status data 202 and the target parameter to be minimised, whilst the binary vector corresponds to the programming data.

The invention provides a control system 100 for improving the flow of a plurality of railway vehicles 101 in a railway station 102 including a plurality of tracks 103.

The system comprises a processor, programmed to perform the steps of the method described in the invention.

According to an embodiment, the system 100 comprises a supervision terminal 10, including at least one display unit 11. The processor, on the basis of the constraining data and/or processing data and/or programming data and/or operating data is programmed to generate image data, intended to be represented on the display unit 11 of the terminal 10, to be able to present a digital representation of the railway station 102.

It should be noted that, by way of example, the reinforcement learning algorithms that can be used and that have shown good performance levels are one or more of the following:

- -q-learning (for a description refer to the contents of the following page https://en.wikipedia.org/wiki/Q-learning), sarsa2 (for a description refer to the contents of the following page:

- augmented random search (ARS) (for a description refer to the contents of the following page

Experiments and tests have shown that, in these circumstances, the best performing algorithm is the augmented random search (ARS), but the others can also be used for the purpose of the invention.

According to an aspect of the invention, the invention provides a computer program comprising instructions for executing the steps of the method described in the invention when launched on the processor of the system 100.