Device and method for the on-line prediction of the driving cycle in an automotive vehicle

Abstract

 The invention relates to a device and method for the on-line prediction of the driving cycle in an automotive vehicle. The method comprises:
- a step of data pre-processing (200):
• receiving the vehicle speed (V_sp);
• receiving traffic information (HTI) of the expected path in a prediction horizon (H);
• obtaining (212) a reference driving cycle (V_pat);
• calculating (208) the deviation (DV_sp) of the speed (V_sp) with respect to the reference driving cycle (V_pat);

- a step of data processing by means of a neural network (202) for recursively obtaining the expected deviations (D*V_sp) for the prediction horizon (H);
- a step of data post-processing (204) which comprises obtaining the estimated speed (V*_sp) for said prediction horizon (H) from the expected deviations (D*V_sp) and the reference driving cycle (V_pat) for the prediction horizon (H).

Description

Field of the Invention

[0001] The present invention is encompassed within the automotive field, and more specifically within the devices and methods for the on-line prediction (while the vehicle is circulating) of the driving cycle of a hybrid vehicle with respect to a preselected prediction horizon. The objective of the invention is to provide the prediction made to the energy management subsystem of the hybrid vehicle so that said vehicle adapts its energy strategy as a function of said prediction, and can thus reduce vehicle consumption as well as optimize the different energy flows found in a hybrid vehicle to increase its energy efficiency, autonomy and reduce CO₂ emissions.

Background of the Invention

[0002] The fact that if the driving cycle (vehicle speed = f (time)) and the terrain slope or gradient cycle (gradient = f (time)) is known beforehand, it would be possible to calculate an optimal energy strategy for the drive system of a hybrid electric vehicle minimizing a cost function made up of terms related to the consumption, emissions, and/or energy efficiency, among others, of the vehicle, is well known.

[0003] To reach the global optimum, there are three drawbacks or barriers to overcome:

1) The driving cycle to be performed by the driver is not known beforehand. Even though the final destination and path to take are known, the driving cycle depends on the driver driving style and on possible disturbances related to the driving environment, such as traffic congestion, meteorological conditions, speed limits due to works, etc.

2) Even though the driving cycle is known beforehand, it is necessary to have a well-modeled vehicle behavior in order to construct the cost function that results from considering the optimization problem.

3) Once the optimization problem is considered, it is necessary to solve it to calculate the global optimum. In this sense, it should be pointed out that a non-linear, non-convex and non-quadratic optimization problem is to be solved, so: (i) there is no explicit or analytical solution for it, and (ii) there are techniques for finding the global optimum, such as Dynamic Programming (DP) but these techniques cannot be computationally treated in systems for on-line real time control. Therefore, certain approaches to problem must be used to solve it with a suitable computational cost. The solution will therefore come close to the global optimum but said global optimum cannot be attained (sub-optimum strategies).

[0004] The present invention focuses on the development of a system or device which contributes to solving the first drawback or barrier related to the prior knowledge of the driving cycle that the vehicle will perform. This system or device therefore obtains the on-line prediction of the future driving cycle (speed* = f* (time)) and the terrain slope or gradient cycle (gradient* = f* (time)) with respect to a preselected prediction horizon, sending this prediction to the energy management system of the hybrid electric vehicle. Therefore, the energy management system of the vehicle could use this prediction in approaching the energy optimization problem and in solving or searching for a solution (energy management/power-energy distribution in the drive system) that is optimum and close to the optimum global solution.

[0005] The model of a driver refers to the representation by means of mathematical formulations or intelligent algorithms of the behavior of the driver of a vehicle, i.e., of the drivers tasks, for analyzing or inferring which actions the driver takes with said vehicle.

[0006] Different driver modeling techniques or algorithms are described in literature reference [1] (Boyraz, Sathyanarayana, & Hansen, 2009). The models were initially linear, being gradually replaced with non-linear, probabilistic models and with intelligent techniques, such as fuzzy logic and neural networks, as described in literature reference [2] (Panou, Bekiaris, & Papakostopoulos, 2007). However, the latest driver modeling trends are heading for a combination of some or all of the aforementioned techniques, referring to this group of models as hybrid. Mealy automata used in [3] (Kiencke & Nielsen, "Road and Driver Models", 2005) for the control logic of the vehicle maneuvers would have to be added in this classification.

[0007] A broader approach relates the model of the driver with the purely dynamic model of the car, as well as with the environment of the driver and his/her vehicle, i.e., the city and other drivers. Different types of driver models can therefore be defined in accordance with the reality which they want to best represent.

[0008] There are many fields of application of said models and the trend in the recent decades is for them to have an increasingly higher repercussion, [2] (Panou, Bekiaris, & Papakostopoulos, 2007). If the different applications are grouped under common umbrellas, there are primarily three trends, to which a fourth is added according to the latest work conducted in this field.

1. Driver behavior in accordance with cognitive and physiological processes

a. Driver behavior analysis

b. Driver behavior inference

c. Driver training and advising

2. Vehicle control

a. Simulation and prototyping

b. Vehicle dynamics

c. Control systems for driving and safety assistance (ABS, ESC, control of traction...)

d. Autonomous driving

3. Traffic simulation

a. Microscopic

b. Macroscopic

4. Energy strategies

[0009] The first trend focuses on the human behavior characteristics of the driver, i.e., the analysis of said behavior, the interpretation of gestures and emotions on one hand and the inference of that behavior in vehicle control, maneuvers and driving strategy. It is obvious to include the Michon hierarchical control model (1985) within this umbrella. The first distinction it makes is to differentiate between external state input-output type models and internal state type models. The other distinction refers to functional models or taxonomy models. Michon asserts that the models are generally bottom-up (internal) and that top-down models are generally non-specifiic and too simplified. His cognitive process type model, the Hierarchical Control Model, divides the task of driving into three coupled and hierarchical levels:

1. The strategic level: planning the path, choosing the route

2. The maneuverability level: relates the driver with the other vehicles

3. The control level: refers to the vehicle control level

[0010] A fourth level would be the purely behavior level, [2] (Panou, Bekiaris, & Papakostopoulos, 2007). Another suitable classification of driver behavior is that which distinguishes between following a desired trajectory and stability in the event of disturbances.

[0011] The second group of applications focuses purely on vehicle control task. This group has a direct correlation with the Michon control level. The control of a vehicle is split in two, longitudinal control (accelerator and brake) and lateral control (steering wheel). Tustin (1947) is considered to be the first author to publish driver model in mathematical form. McRuer & Krendel, Ragazzini and Jackson ([4] Abe, 2009) later joined Tustin in making contributions of interest.

[0012] Although in principle longitudinal and lateral control were controlled independently, recent contributions relate both controls because one affects the other ([1] Boyraz, Sathyanarayana, & Hansen, 2009). These controls are studied together with the different maneuvers of the vehicle. The three main maneuvers are lane keeping, lane change and speed control according to traffic signals and tracking a vehicle. For a comprehensive critical summary of these techniques see [5] (Khodayari, Ghaffari, Ameli, & Flahatgar, 2010). In addition to these controls specific to each maneuver a decision logic is provided for establishing which maneuver is suited to each situation. Reference [3] (Kiencke & Nielsen, Road and Driver Models, 2005) proposes a Mealy automaton.

[0013] The third type of models focuses on traffic simulation both microscopically (individual behavior with respect to the traffic) and macroscopically (large environments with several drivers) ([6] Fernandez, 2010).

[0014] The fourth type of models refers to those used in searching for an optimum energy management system for hybrid vehicles. The importance of the driver model in these applications stems from the fact that its behavior considerably affects the energy distribution.

[0015] A new driver model is specifically developed in literature reference [7] (Froberg & Nielsen, 2008) for dynamic inverse simulations for the purpose of optimizing the simulation times used in searching for energy and drive strategies. Its objective is driving cycle simulation efficiency, not real time implementation.

[0016] However, in addition to the effect of the driver, the type of thoroughfare, the traffic situation, the operating mode and the driving trend also have a considerable effect ([8] Murphey Y., 2008). The key to this problem is predicting driver behavior and the driver environment to thus determine the optimum operating mode. Only recently have research efforts started to focus on this track. The algorithms used are partially based on intelligent techniques.

[0017] A relevant paper is that citied in reference [9] (Langari & Won, 2005). This paper presents a system for identifying the driving cycle by means of an LQV neural network and a fuzzy logic system for the purpose of improving energy management (IEMA). By analyzing the current driving cycle, it tries to identify it by comparing it with 9 previously selected driving cycles of different thoroughfares. Therefore, the prediction does not take into account the future environment data, but rather it estimates them from typical cycles.

[0018] In a very similar manner, reference [10] (Murphey, et al., 2008) develops its own algorithm and said algorithm is applied to a conventional vehicle, achieving fuel reductions of up to 2.68% in real time. The authors emphasize that the improvement with off-line DP (Dynamic Programming) algorithm is 2.81%. Its IPC system is based on neural networks for the prediction of the current driving environment. It is then analyzed with typical driving cycles for its identification. An evolution of this same algorithm can be found in [11] (Park, et al., 2009). Like in the preceding reference, the future environment data is not taken into account.

[0019] Another similar algorithm is that shown in reference [12] (Huang, Tan, & He, 2010). The authors first perform a statistical analysis of the driving cycle sampled in real time and then predict the driving conditions by means of an SVM (support vector machine) and a neural network. It has a 95% precision. However, the entire prediction is based on data collected in real time, so it does not include future environment data.

[0020] Another approach to the problem is by means of driving pattern recognition by means of fuzzy logic and the subsequent prediction by means of Hidden Markov Models ([13] Montazeri- Gh, Ahmadi, & Asadi, 2008). The optimum is not assured because fuzzy logic is used.

[0021] The success of the control strategy largely depends on the quality of that prediction as well as the length of said prediction ([14] Koot, Kessels, Jager, Heemels, den, & Steinbuch, 2005). The algorithm used must be optimum, although given the complexity of the problem a quasi optimum is achieved. To include the greatest possible amount of driver environment information, data about the road state and traffic congestion and thoroughfare type and speed limit must be obtained. With the aid of modern GPS type navigation systems it would be possible to obtain this data. This technology can result in a greater efficiency and precision in predicting driver behavior and thereby the intelligent drive system behavior, which is fundamental in the global efficiency of electric hybrid vehicles.

Literature

[0022] 

[1] Boyraz, P., Sathyanarayana, A., & Hansen, J. H. (2009). "Driver behavior modeling using hybrid dynamic systems for 'driver-aware' active vehicle safety. (ESV) Enhanced Safety for Vehicles", -, 13-15.

[2] Panou, M., Bekiaris, E., & Papakostopoulos, V. (2007). Modelling Driver Behaviour in European Union and International Projects. In Modelling Driver Behaviour in Automotive Environments (pp. 3-25). Springer London.

[3] Kiencke, U., & Nielsen, L. (2005). Road and Driver Models. In Automotive Control Systems (pp. 425-464). Springer Berlin Heidelberg.

[4] Abe, M. (2009). Vehicle Handling Dynamics. El Sevier.

[5] Khodayari, A., Ghaffari, A., Ameli, S., & Flahatgar, J. (2010). A historical review on lateral and longitudinal control of autonomous vehicle motions. (pp. 421-429).

[6] Fernandez, A. (2010). Simulación del comportamiento de los conductores mediante agentes inteligentes. Universidad Complutense Madrid.

[7] Froberg, A., & Nielsen, L. (2008). Efficient Drive Cycle Simulation. Vehicular Technology, IEEE Transactions on, 57, 1442-1453.

[8] Murphey, Y. (2008). Intelligent Vehicle Power Management - An Overview. In Computational Intelligence in Automotive Applications (Vol: 132, pp. 223-251). Springer Berlin / Heidelberg.

[9] Langari, R., & Won, J.-S. (2005). Intelligent energy management agent for a parallel hybrid vehicle-part I: system architecture and design of the driving situation identification process. Vehicular Technology, IEEE Transactions on, 54, 925-934.

[10] Murphey, Y., Chen, Z. H., Kiliaris, L., Park, J., Kuang, M., Masrur, A., et al. (2008). Neural learning of driving environment prediction for vehicle power management., (pp. 3755-3761).

[11] Park, J., Chen, Z., Kiliaris, L., Kuang, M., Masrur, M., Phillips, A., et al. (2009). Intelligent Vehicle Power Control Based on Machine Learning of Optimal Control Parameters and Prediction of Road Type and Traffic Congestion. Vehicular Technology, IEEE Transactions on, 58, 4741-4756.

[12] Huang, X., Tan, Y., & He, X. (2010). An Intelligent Multifeature Statistical Approach for the Discrimination of Driving Conditions of a Hybrid Electric Vehicle. Intelligent Transportation Systems, IEEE Transactions on, PP, 1-13.

[13] Montazeri-Gh, M., Ahmadi, A., & Asadi, M. (2008). Driving condition recognition for genetic fuzzy HEV Control. (pp. 65-70).

[14] Koot, M., Kessels, J., Jager, B. d., Heemels, W., den, P. v., & Steinbuch, M. (2005). Energy management strategies for vehicular electric power systems. Vehicular Technology, IEEE Transactions on, 54, 771-782.

Description of the Invention

[0023] The present invention consists of a device and a method for the on-line prediction (while the vehicle is circulating) of the driving cycle in a vehicle with respect to a preselected prediction horizon. The proposed device is based on a prediction strategy made up of a step of pre-processing the inputs received, an artificial neural network (ANN), and a step of post-processing for obtaining the prediction. The strategy is supported on three main information sources:

• Examples/data used for training the artificial neural network.
• Traffic database (traffic signals according to path to take).
• Second generation navigation system with real time traffic information. Use of the information (traffic events and state) facilitating such systems in an anticipated manner.

[0024] The device predicts the future driving cycle (speed vs. time, and path gradient vs. time) while the vehicle is circulating. For that purpose, the device uses certain information (measurements) of the driving cycle performed in the recent past, as well as certain information (from the new navigation systems which are connected with traffic management systems in real time, second generation systems)) of an advanced nature related to future traffic events that will occur in the path that it being taken.

[0025] The main objective of this device and method is to provide the prediction made to the energy management subsystem of the hybrid vehicle so that said vehicle adapts its energy strategy as a function of said prediction, and can thus reduce vehicle consumption as well as optimize the different energy flows found in a hybrid vehicle to increase its autonomy and reduce CO₂ emissions.

[0026] The proposed device has a communication channel for connecting to the vehicle second generation navigation system, and another communication channel for connecting to the subsystem (Electronic Control Unit, ECU) which performs energy management functions in a hybrid vehicle.

[0027] The proposed device/method receives the current vehicle speed measurement, as well as navigation subsystem information related to information about the path that is being taken (speed limits, road slopes,...), as well as real time traffic events information about events that are occurring in future points where the vehicle will be circulating (jams, holdups,...).

[0028] This information (speed measurement and navigation system information) is preprocessed and used as input in a neural network with a specific topology which obtains mainly two outputs. The first output corresponds with the prediction of the driving cycle (speed vs. time*) made with respect to a selected prediction horizon (usually in Km), whereas the second output corresponds with the prediction of the slope cycle (slope vs. time*) with respect to the selected horizon. The proposed device and method further obtains a third output related to the user driving style. To obtain this third output a fuzzy inference system is used using specific expert rules which are based on specific parameters which are being obtained using the driving cycle measurements from the recent past.

[0029] Finally, it must be pointed out that the proposed prediction algorithm and technique (based on neural network, fuzzy inference system and steps of pre-processing and post-processing of the information) works on-line while the vehicle is circulating, calculating and obtaining new predictions (prediction update) in each sampling instant selected for it.

[0030] The neural network used tries to obtain the non-linear function which best approximates the current vehicle-driver circulation. For that purpose, this network gradually learns the vehicle behavior and the different driving modes of the usual driver by means of the driving cycles that are ultimately being performed. Therefore, it relates to a device/method with self-learning capability.

[0031] As far as differences with respect to other similar systems, it can be pointed out that this device/method does not try to characterize or classify the driving cycle which is being performed (like most of the reviewed systems/techniques do) according to pattern cycles, and take actions based on this classification as a function of a rule-based system, but rather the proposed device obtains the prediction using information taken from the recent driving past and from future traffic events information, thus being integrated with navigation systems using new technologies (mobile networks, traffic management, connection to management servers and traffic supervision, etc.) for the purpose of constructing on-line (while circulating) the most probable driving cycle that the vehicle-driver will follow in future kilometers of the path that is being taken. Furthermore, the device/method proposed is capable of making the prediction without knowing beforehand the final destination of the path, although it is more precise and accurate and offers enormous potential if the route and/or final destination is known beforehand: indicated in the navigation system by the driver or even recognizable by the system itself (paths usually taken). Therefore, if the final destination and route that is being taken are known beforehand, the selected prediction horizon can be very large, obtaining a high prediction precision. However, if the final destination and route are not known, the selected prediction horizon must be smaller if prediction precision is to be maintained, or otherwise (large prediction horizons) the prediction precision could be affected because the system or device has to opt for (or has to guess) one route from among all those possible routes that the driver will follow based on statistical data and probabilities. It must be pointed out that even though the destination/route is not known beforehand, the device will continue operating though with a certain penalization in the precision. Furthermore, and because the predictions are performed on-line (while the vehicle circulates), such predictions will be renewed in every sampling instant and therefore the route decisions that the driver takes are taken into account in the "recalculation" of the predictions in each sampling instant.

[0032] The method for the on-line prediction of the driving cycle in an automotive vehicle comprises:

• a step of data pre-processing which in turn comprises:
  • receiving the vehicle speed;
  • receiving traffic information corresponding to the expected path for the vehicle within at least one prediction horizon considered;
  • obtaining a reference driving cycle corresponding to the expected path within at least said prediction horizon from the traffic information received;
  • calculating the deviation of the vehicle speed with respect to the reference driving cycle;
• a step of data processing by means of a neural network, which comprises recursively obtaining the expected deviations for the prediction horizon, using for that purpose the deviations of speed previously calculated and corresponding to the recent past in a delay sample number as well as information relating to the reference driving cycle containing information belonging both to the recent past in a delay sample number and to the near future in a future sample number as inputs of the neural network;
• a step of data post-processing which comprises obtaining, the estimated speed for said prediction horizon from the expected deviations and the reference driving cycle for the prediction horizon.

[0033] The step of data pre-processing can comprise receiving traffic events information corresponding to the expected path for the vehicle within at least the prediction horizon, and where said traffic events information received is used also for obtaining the reference driving cycle.

[0034] The neural network is preferably a previously trained recurrent dynamic neural network with NARX topology.

[0035] In a preferred embodiment, the vehicle speed is sampled in the step of data pre-processing according to a specific sampling time, and obtaining the reference pattern cycle and the calculation of the deviation of the vehicle speed with respect to the reference driving cycle are performed for each sampling time.

[0036] The information relating to the reference driving cycle can comprise a pattern speed moved forward a future sample number, which is equivalent to the vision distance of the driver and the anticipation of the driver with respect to future traffic situation changes.

[0037] The traffic information can additionally include at least one of the following pieces of information:

• the speed limits;
• information of the type of thoroughfare;
• the road slopes;
• the traffic signals of the expected path.

[0038] The traffic events information can include information relating to at least one of the following:

• traffic state;
• speed limits due to road works;
• visibility conditions;
• road surface conditions.

[0039] The traffic information and the traffic events information are preferably received within the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon.

[0040] The method can comprise obtaining the driving style of the driver of the vehicle according to calculations depending on a parameter relating to the driving style calculation mode selected, where the calculation modes are based on at least one of the following:

• calculation based on Fourier transform of a vector formed by the vehicle speed values corresponding to the recent past;
• calculation based on the mean speed variation over a period of time;
• calculation based on driver reaction times.

[0041] Another object of the present invention is a device for the on-line prediction of the driving cycle in an automotive vehicle, comprising:

• communication means configured for receiving the vehicle speed and for receiving traffic information corresponding to the expected path for the vehicle within at least one prediction horizon considered from a navigation system;
• data processing means configured for:
  • obtaining a reference driving cycle corresponding to the expected path within at least said prediction horizon from the traffic information received by the communication means;
  • calculating the deviation of the vehicle speed with respect to the reference driving cycle;
  • recursively obtaining the expected deviations for the prediction horizon by means of a neural network, using for that purpose the deviations of speed previously calculated and corresponding to the recent past in a delay sample number as well as information relating to the reference driving cycle containing information belonging both to the recent past in a delay sample number and to the near future in a future sample number as inputs of the neural network;
  • obtaining the estimated speed for said prediction horizon from the expected deviations and the reference driving cycle for the prediction horizon.

[0042] The communication means can additionally be configured for receiving traffic events information corresponding to the expected path for the vehicle within at least the prediction horizon from the navigation system, and where the data processing means are configured for obtaining the reference driving cycle also using said traffic events information received by the communication means.

[0043] The data processing means are preferably configured for sampling the vehicle speed according to a specific sampling time and for obtaining the reference pattern cycle and the calculation of the deviation of the vehicle speed with respect to the reference driving cycle for each sampling time.

[0044] The device can also comprise the actual navigation module.

[0045] The data processing means can be configured for performing the prediction calculation while the vehicle is circulating and every time the vehicle advances by a selected distance by means of a parameter.

Brief Description of the Drawings

[0046] A set of drawings which aid in better understanding the invention and which expressly relate to an embodiment of said invention presented as a non-limiting example thereof is very briefly described below.

Figure 1 shows the structure of the system for the prediction of a driving cycle proposed in the present invention.

Figure 2 shows the strategy for the prediction of the driving cycle.

Figure 3 depicts a real driving cycle and a pattern cycle as a function of the kilometer marker.

Figure 4 shows a structure of the proposed NARX neural network.

Figure 5 shows a specific embodiment of the NARX neural network for an embodiment of the present invention with its inputs and outputs.

Figure 6 shows a diagram with the components of the device for prediction object of the present invention.

Detailed Description of the Invention

[0047] The structure of the proposed device for the prediction of the driving cycle is presented in Figure 1 as a black box with inputs and outputs.

[0048] As can be observed, the device for the prediction of the driving cycle 100 has the instantaneous vehicle speed measurement (in Km/h) and two specific inputs referring to traffic information obtained by means of a navigation system 104, Traffic Information (HTI, Horizon Traffic Information) and Traffic Events Information (HTEI, Horizon Traffic Events Information), as inputs. The Traffic Information (HTI) input contains the speed limits, road slopes and traffic signals in the prediction horizon considered. The Traffic Events Information (HTEI) input contains information such as state/flow of traffic, works, visibility and road surface conditions. The prediction horizon (H) is the driving cycle prediction interval.

[0049] The input parameters of the device for the prediction of the driving cycle 100 are explained in detail below:

• Ignition (IG): Activation of the device for the prediction of the driving cycle 100, performed through the ignition system 102 or the actual energy management system (EMS) of the vehicle 108. This input is used to synchronize the system with the application that will use the predictions made. When this input is activated at "1", the rising edge is used to set the time variable to 0 (t=0) and the prediction strategy and calculation of the outputs is initiated.
• Vehicle Speed (V_sp): Instantaneous vehicle speed (usually in Km/h). This input (measurement) could be obtained from either the vehicle navigation system 104 (as shown in Figure 1) or from the vehicle control unit by means of the communications bus thereof.
• Traffic information (Horizon Traffic Information, HTI): Structure type input obtained from the vehicle navigation system 104 with real time traffic information containing the following vectors:

  * Speed_Limits (Horizon Speed Limits, HTI_HSL) vector: The vector containing the speed limits in ideal circulation conditions existing in the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon (usually in Km). Therefore they are the speed limits defined by traffic (traffic signals corresponding to speed limits) existing in the path that is being taken. The size of this vector depends on the desired resolution (H_Resol parameter (usually in Km)) and on the size of the selected prediction horizon H.

[0050] The size will therefore be determined by: Size = H / H_Resol. It should be indicated that instead of said Speed_Limits vector, it could be replaced with any other vector providing information from which the device can deduce speed limits; for example, the type of thoroughfare (highway, road with a wide shoulder, etc.) when there is no higher limitation.

* Road_Slopes (Horizon Road Slopes, HTI_HRS) vector: The vector containing the road slope % value existing in the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon. The size of this vector depends on the desired resolution (H_Resol parameter) and on the size of the selected prediction horizon H. The size will therefore be determined by: Size = H / H_Resol.

* Stop_Signals (Horizon Stop Signals, HTI_HSS) vector: The vector containing future stop signals, yield signals, stoplights and road toll booths existing within the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon. The size of this vector depends on the desired resolution (H_Resol parameter) and on the size of the selected prediction horizon H. The size will therefore be determined by: Size = H / H_Resol. By way of example, the values that the vector contains could be:

0 => There is no signal

1 => Stop signal

2 => Yield signal

3 => Toll booth

4 => Stoplight

[0051] As can be observed, it is information that is known beforehand according to the path that is being taken. Navigation systems have this information in their databases and could provide it in an anticipated manner with respect to the selected prediction horizon.

• Traffic Events Information (Horizon Traffic Events Information, HTEI): Structure type input containing the following vectors:

  * Traffic_State (Horizon Traffic State, HTEI_HTS) vector: The vector containing the traffic status within the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon. The size of this vector depends on the desired resolution (H_Resol parameter) and on the size of the selected prediction horizon H. The size will therefore be determined by: Size = H / H_Resol. The values that this vector contains depend on the traffic state within the selected prediction horizon which will be determined by the navigation system and its real time traffic information. Each vector element could have different values such as the following:

  0 => Very light traffic.

  1 => Light traffic.

  2 => Heavy traffic.

  3 => Very heavy traffic.

  4 => Stopped traffic.

  * Road Works (Horizon Road Works, HTEI_HRW) vector: The vector containing the speed limits due to road works existing within the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon. The size of this vector depends on the desired resolution (H_Resol parameter) and on the size of the selected prediction horizon H. The size will therefore be determined by: Size = H / H_Resol. The values that this vector contains depend on the works in the road within the selected prediction horizon which will be determined by the navigation system and its real time traffic information.

  * Visibility_Conditions (Horizon Visibility, HTEI_HV) variable: The variable containing the status of the visibility conditions (effect of fog, rain or snow) within the selected horizon. By way of example, it could have the following values:

  0 => Good visibility.

  1 => Average visibility.

  2 => Poor visibility.

  3 => Very poor visibility.

  * Road_Conditions (Horizon Road Condition, HTEI_HRC) variable: The variable containing the status of the road road surface (effect of ice, water, etc.) within the selected horizon. By way of example, it could have the following values:

  0 => Road surface in good conditions.

  1 => Somewhat slippery road surface.

  2 => Slippery road surface.

  3 => Very slippery road surface.

[0052] As can be observed, in this case these variables are not known beforehand and have a dynamic character. The navigation systems to be used must have the characteristic of being able to obtain real time traffic information and events. Some types of models which can obtain this type of real time information are starting to be sold today. For that purpose, the devices are either connected to traffic management systems via communications (RDS, 802.11x, etc.) or obtain the information by creating communication networks the users of which are onboard vehicle navigation systems. These vehicles/users share information with a server which infers the traffic state based on the speed and position measurements it receives from the different vehicles forming the network. Today there are navigation system models which share information from several million users/vehicles in specific geographic areas or regions.

[0053] The device for the prediction of the driving cycle 100 uses the following parameters:

• Sampling time (ST): Selected sampling time expressed in seconds. Real type variable. The device for the prediction of the driving cycle 100 samples the instantaneous vehicle speed V_sp input according to the value introduced in this parameter.
• Kalman Filter Level (KFL): The device for the prediction of the driving cycle 100 filters the instantaneous vehicle speed V_sp input in real time according to the value introduced (between 0 and 4) in this parameter.
• Prediction Horizon (H): Desired prediction horizon, usually expressed in Km. The prediction with respect to the selected horizon: p -> p + H will be performed from the current vehicle position (p).
• Horizon Resolution (H_Resol): Desired resolution for the selected horizon, usually expressed in Km. For example, for a prediction horizon H = 10 Km, if an H_Resol value = 0.01 Km is selected, the size of the prediction horizon vector will be of 1000 elements. This parameter further represents how many kilometers the vehicle advances the prediction is performed/recalculated, for example if H_Resol = 0.05 Km, the prediction will be performed/recalculated every 50 m of vehicle advancement.
• Delay Sample Number (DSN): Number of samples selected to form the recent past. For example, a value of 40 means that the distance considered as the recent past is 40xH_Resol (Km), i.e. the recent past would be the last 40xH_Resol Kilometers.
• Future Sample Number (FSN): Number of samples selected to form the near future. For example, a value of 2 means that the distance considered as the near future is 2xH_Resol (Km), i.e., the near future will be the next 2xH_Resol kilometers. This parameter is usually chosen so that the distance considered as the near future is equivalent to the vision distance, anticipation and reaction of the driver. For example, an FSN value making the near future distance 0.2 Km means that the range of vision and therefore of anticipation of the driver with respect to future speed limit (traffic signals) changes is 200 meters.
• Driving Style Calculation Mode (DSCM): Driver driving style calculation mode. The manner in which the device calculates the driving style is selected by means of this parameter. The calculation modes can be: calculation based on the Fourier transform (mean value and value of the first fundamental harmonic or signal) (DSCM = 1), calculation based on the mean speed variation over a period of time (DSCM = 2), or calculation based on driver reaction times (DSCM = 3).

[0054] The device for the prediction of the driving cycle 100 obtains the following outputs, shown in Figure 1, which are provided to an external unit 108, which can be the Energy Management System (EMS) of the vehicle or any other third party application:

• Estimated speed (V*_sp): The prediction of the speed with respect to the horizon H is performed in space and over time. Specifically:
  • V*_sp [p, p+H)]: the speed in the spatial interval [p, p+H)] is estimated, p being the current vehicle position and H the selected prediction horizon.
  • V*_sp [t, t+T_H]: the speed in the time interval [t, t+T_H] is estimated, t being the current instant in time and T_H (s) the estimated time in which the vehicle will reach the prediction horizon.
• Estimated Road Slope (S_est): the prediction of the road slope with respect to the horizon H is performed in space and over time. Specifically:
  • S*[p, p+H]: the slope in the spatial interval [p, p+H)] is estimated, p being the current vehicle position and.H the selected prediction horizon.
  • S* [t, t+T_H]: the slope in the time interval [t, t+T_H] is estimated, t being the current instant in time and T_H (s) the estimated time in which the vehicle will reach the prediction horizon.
• Driving Style (DS): The driving style of the driver of the vehicle is obtained.

[0055] The strategy, algorithms and technique used for obtaining each of said outputs of the device for the prediction of the driving cycle 100 is explained below:

1. Prediction of estimated speeds (outputs V*_sp [p, p+H] and V*_sp [t, t+T_H])

[0056] The created prediction strategy is based on using a previously trained Artificial Neural Network (ANN) with NARX (nonlinear autoregressive network with exogenous inputs) topology. This strategy is completed with a series of pre- and post-processing functions of both the inputs and the outputs of this ANN-NARX. The objective of the ANN-NARX is to learn the behavior and driving mode of the driver-vehicle combination by evaluating the deviations of vehicle speed with respect to a reference driving cycle or pattern corresponding to the path that is being taken.

[0057] The reference driving cycle or pattern for the future horizon is dynamically constructed in each sampling time as a function of the information received through the Traffic Information (HTI) and Traffic Events Information (HTEI) inputs. The ANN-NARX also obtains the expected deviations of speed (with respect to the reference cycle) for the future horizon in each sampling instant using the reference driving cycle (pattern cycle) and the deviations of speed occurring in the recent past of the path for that purpose. Therefore, the real expected driving cycle with respect to the selected prediction horizon can finally be obtained by using the prediction of the expected deviations of speed and the expected reference cycle.

[0058] The block diagram corresponding to the strategy used by the device for the prediction of the driving cycle is shown in Figure 2. The three steps of the strategy, the pre-processing (200), the artificial neural network (202) and the post-processing (204), are explained below in detail. The pre-processing (200), the artificial neural network (202) and the post-processing (204) are done by means of data processing, using for example a system or device based on a microcontroller or microprocessor supported by a set of memory elements and input/output and communication ports.

[0059] In the pre-processing (200), the first function that is performed consists of applying a real time filter 206 on the Vehicle Speed (V_sp) input variable. In a preferred embodiment, the type of filter applied is a real time Kalman Filter. The second function in the pre-processing (200) consists of performing a domain transformation 207 on the V_spkf variable obtained as output after applying the filter 206. A transformation from the time domain to distance domain (kilometer marker) is performed by means of this second function 207 obtaining the internal Kilometer_Marker (PKm) variable. This variable is obtained by means of numerical integration of the speed V_spkf. Therefore, this function gradually generates a two column vector [V_spf(i), PKm(i)] where each row "i" is calculated according to the algorithm presented in (1) in each sampling time "k". It can be observed that the PKm(i) vector is always increasing and further has a constant sampling "i" which corresponds with a distance of H_Resol. The V_spf(i) variable represents the vehicle speed corresponding to each kilometer marker PKm(i).

[0060] The next function that is performed in this step is the calculation of the deviation vector 208, which consists of calculating' the deviation existing between the vehicle speed V_spf(i) and the pattern speed or reference driving cycle V_pat(i) (V_pat(i) being the speed allowed for said kilometer marker i) for each kilometer marker PKm(i). The calculation for constructing the deviation vector DV_sp(i) is shown in the algorithm (2) and is performed as the PKm(i) and V_spf(i) vectors are being constructed.

[0061] The pattern speed vector V_pat(i) is constructed using the available traffic information from the navigation system (Horizon Traffic. Information, HTI, input). If the path to take (final destination and route to follow) is known, the traffic signals corresponding to speed limits existing in each kilometer marker of said route or path to take can be known. An example can be observed in Figure 3, in which a real driving cycle performed on a specific path marked in blue the vehicle speed V_spf(i) and the corresponding pattern speed vector V_pat(i) marked in red, has been represented by way of example. It can be observed that both vectors are represented with respect to the Kilometer_Marker PKm(i) vector. The cycle represented by way of example in Figure 3 has been performed with a vehicle that was equipped with an onboard data acquisition system. If the path to take is not known, said path must be estimated by probabilistic methods in order to construct the pattern speed vector for the future horizon. Logically in this case, the higher the selected prediction horizon, the error in obtaining the pattern speed vector and therefore in the final prediction made could be penalized.

[0062] The prior knowledge of the path to take or, in its absence, the estimation thereof, allows constructing the pattern speed vector V_pat(i) in each sampling time in the step of constructing a pattern cycle 212 of Figure 2 using the Traffic Information (HTI) input. Once the vector is constructed, said vector could vary according to different events that can occur as a result of the different traffic conditions which exist in said path and are unpredictable by nature. Therefore, this pattern speed vector V_pat(i) could gradually change or be adapted while the vehicle is circulating (on-line), using the real time traffic information received through the navigation system 104 by means of the Traffic Events Information (HTEI) input for said adaptation.

[0063] Finally, in this step of pre-processing 200, the vectors NN_DV_sp(i), and NN_V_pat(i), which are the inputs of the artificial neural network 202, are generated. The NN_DV_sp(i) vector is generated in step 209 and contains the last DSN samples of the DV_sp(i) vector, where DSN (parameter of the device for the prediction of the driving cycle 100) depicts the number of samples defining the size of the NN_DV_sp(i) vector, or in other words, the size of the recent past. Therefore, the NN_DV_sp(i) vector contains the last DSN values of the DV_sp(i) vector, see algorithm (3).

[0064] The NN_V_pat(i) vector is generated in step 210 and on one hand contains the last DSN samples of the vector V_pat(i), where DSN represents the number of samples defining the size of the recent past, and on the other hand the future FSN samples of the V_pat(i) vector, where FSN (parameter of the device for the prediction of the driving cycle 100) represents the number of samples defining the size of the near future. Therefore, the NN_V_pat(i) vector contains the last DSN values of the V_pat(i) vector, and the future FSN values of the V_pat(i) vector, see algorithm (4).

[0065] In summary, the functions performed in this step of pre-processing 200 are aimed at constructing the vectors NN_DV_sp(i), and NN_V_pat(i) which are the inputs to the artificial neural network 202 created and used for performing the prediction of the driving cycle.

[0066] The topology of the artificial neural network 202 selected for performing the prediction is within recurrent dynamic neural networks and is referred to as NARX (nonlinear autoregressive network with exogenous inputs). The NARX model is based on the linear ARX model which is commonly used in time series analysis and prediction. The equation defining the non-linear NARX model is shown in (5), where y represents the output variable and x₁, x₂, ..., x_n, represent the possible inputs of the model or system to be modeled, f represents a possible non-linear function, and NumDelaysY, NumDelaysX_1..n., represent the number of previous samples that are taken into account for calculating the output of the non-linear function.

[0067] A non-linear NARX model can be implemented through a feed-forward neural network approaching the non-linear function f. Figure 4 shows a diagram of the structure of the proposed neural network 202 in which the inputs 400, the output 408 and the input layer 402, hidden layer 404 and output layer 406 which are formed by neurons can be observed. As can be observed, the inputs and the output correspond with those present in (5).

[0068] Each input in a neural network corresponds to a series of parameters (weights and bias 410) joining each input neuron with the corresponding neurons forming the hidden neuron layer 404. Therefore, the value of each neuron belonging to the hidden layer 404 is calculated by applying a usually non-linear function to the sum of the product of the input neurons by their corresponding weights added to the biases. This operation is gradually transmitted to the neurons forming the output layer 406, where the value of each output is obtained by means of applying a linear or non-linear function to the sum of the product in feed-forward direction.

[0069] Given several examples/tests conducted where the input and output data of the non-linear function which is to be approximated by an NARX type neural network have been recorded, the process of training the neural network defined consists of obtaining the parameters thereof (weights and bias 410 of all the connections) which lead to obtaining the outputs of the examples with a minimal error for the inputs of the examples. Therefore, the neural network "learns" from these examples, acquiring the property of "generalizing" when other sequences and different inputs occur in the network. The success of training the network depends on the training algorithm used, on the number of layers and neurons selected, and especially on the examples used for training it, which must have enough information for the network to acquire the property of generalizing and not the property of over-learning the examples used.

[0070] For the case at hand in relation to the prediction of the driving cycle, the output to be estimated (network output) is the deviation of speed in the next kilometer marker, D*V_sp (k+1). The variable NN_DV_sp together with its corresponding previous DSN samples, the variable V_pat together with its corresponding previous DSN samples and subsequent FSN samples, are inputs, see Figure 5.

[0071] The network calculates the output in the instant k+1, but since in this case the prediction is to be made with respect to a horizon H, this calculation can be repeated n times in a recursive manner and the prediction of the expected deviation of speed (D*V_sp) with respect to the selected prediction horizon H can therefore be calculated. Logically, to perform this prediction with respect to horizon H, the actual estimated variables of the output of the network (D*V_sp) are needed, as represented by the dotted line in Figure 5. The basic algorithm for performing the prediction with respect to the horizon H is shown in (6), where the input vectors in the network together with the corresponding samples are constructed in the preparation_inputs_NARX() sub-function, see Figures 2 and 5. To construct the NN_DV_sp input vector, it is necessary to use the actual estimated output of the network D*V_sp when the neural network is being applied during the prediction horizon H, because the actual values of the variable DV_sp are still unknown. Therefore, vectors V_pat, DV_sp, and D*V_sp are necessary to prepare the inputs of the network, as shown in (6).

[0072] After performing several examples and trainings both with data from real driving cycles and with data obtained by means of a virtual driving simulator in scenarios where "non-city" driving predominates with respect to "city" driving in the path, the parameters and the structure of the network which best approximate these deviations is the following:

H_Resol = 0.05 Km

DSN = 40 delays (equivalent to a recent past of 2 Km)

FSN = 4 (equivalent to a near future of 0.2 Km)

Number of neurons in the hidden layer = 20

Functions used in hidden layer neurons = Tangent sigmoid.

Function used in output layer neuron = Linear (purelin).

[0073] As shown above, the NARX neural network 202 obtains the estimation of the deviation of speed expected for the selected horizon, from i to i+H/H_Reso/, as output for each kilometer marker i (PKm(i)).

[0074] The purpose of the functions performed in the step of post-processing 204 is to finally obtain the outputs:

• V*_sp [i, i+H]: Expected vehicle speed vector for the future horizon in the kilometer marker domain. Speed = f(PKm)
• V*_sp [t, t+T_H]: Expected vehicle speed vector for the future horizon with respect to the time domain. Speed = f (time)

[0075] Obtaining the vector speed V*_sp [i, i+H] in the domain corresponding to the Kilometer_Marker (PKm) is simple and direct. Only the points of the pattern speed vector corresponding to the selected horizon need to be used. The calculation thereof is shown in (7), subtracting the estimation of the deviation of speed expected for the selected horizon (DV*_sp).

[0076] Obtaining the vector of speed V*_sp [t, t+T_H] in the time domain is somewhat more complex because it requires performing the transformation 222 from the "Kilometer Marker" domain to the time domain. A distance equivalent to H_Resol is traveled in each spacing i, so by knowing the speed in said instant the increase in time that had to be used to travel said distance at said speed in each spacing i can be estimated. Equation (8) shows the calculation that is performed.

[0077] Therefore, the time variable starts with the current instant in time and the increases in time in each spacing i in the prediction interval are calculated and accumulated according to (9).

2. Prediction of estimated slopes (outputs S* [i, i+H] and S*[t, t+T_H])

[0078] Obtaining the S* [i, i+H] vector, containing the road slopes with respect to the prediction horizon from the current kilometer marker is direct because the road slope at all the kilometer markers of the path that is being taken or is expected to be taken (S_road vector) is previously known. Assuming that these slopes are stored in the Road Slopes (HTI_HRS) vector, the algorithm shown in (10) is followed.

[0079] For obtaining the S* [t, t+T_H] vector, corresponding to the slopes with respect to the prediction horizon from the current instant in time, it is necessary to also perform a domain transformation from the "Kilometer Marker" domain to the time domain. A process similar to the one explained above in the step of post-processing 204 is followed for that purpose. Therefore, to each point S(i) of the interval corresponding to the prediction horizon corresponds an instant in time, time*(i) according to (11).

3. Driving style (DS output)

[0080] The driving style is obtained in different manners according to the selection made by means of the DSCM (Driving Style Calculation Mode) parameter. The manners in which it is obtained depending on the selection made are described below:

- Calculation based on Fourier transform: DSCM = 1

[0081] This method seeks to observe the speed variations being caused by the driver with respect to the vehicle as well as the frequencies thereof. It seems logical to think that, for example, in a section where the speed limit is 80 Km/h (pattern speed), if the vehicle speed is higher and oscillating around this limit, the probable conclusion would be that the driver is in a hurry and is driving aggressively. The Fourier transform applied to the speed signal of the vehicle offers the possibility of calculating the mean or continuous value of the signal (DC value) as well as the amplitudes of the main harmonic and of other orders. By relating these values with the pattern speed signal, conclusions as to the driving mode can be drawn. Therefore, the calculation process consists of applying the Fourier transform to the vehicle speed signal with respect to an interval corresponding to the recent past, obtaining the mean value and amplitude magnitudes of the first harmonic (fundamental signal) and relating them with their corresponding magnitudes in the pattern speed signal to see the existing variation. This process is shown in (12), where D represents the desired number of previous samples (recent past) that are evaluated in the algorithm, DC represents the continuous value of the signal after performing the Fast Fourier transform, and A1 represents the amplitude of the first harmonic (fundamental signal) after the transform. The use of the index i in the algorithm (Kilometer Marker domain) should be pointed out.

- Calculation based on the mean speed variation over a period of time: DSCM = 2

[0082] This method seeks to observe the mean vehicle speed variation in a predetermined time interval. This measurement could also indicate the degree of aggressiveness in driving. The process for obtaining it is presented in (13). As can be observed, this time the index k (time domain) is used, obtaining the mean speed variations with respect to the mean speed value in the time interval corresponding to D x ST, ST being the sampling time selected by means of the corresponding parameter.

- Calculation based on driver reaction times: DSCM = 3

[0083] This method is based on obtaining the driver anticipation time in the event of a future speed limit change. For example, a driver who starts to accelerate 100 meters before the signal of traffic corresponding to a speed limit change (to a higher value) is most likely driving aggressively. Also, a driver who starts to brake at a kilometer marker that is higher than the corresponding kilometer marker where a traffic signal indicating a speed limit change (to a lower value) is located is assumed to be driving aggressively. The measurement of these anticipation and delay values in the recent past combined with the deviation of speed occurring in the steady systems (at a constant speed, as in the preceding methods) could provide a more accurate driving style.

[0084] Figure 6 shows in a non-limiting manner the components of a device for the prediction of the driving cycle 100 which performs the steps of the method described above.

[0085] The device comprises first communication means 600 for receiving the vehicle speed and traffic information, usually from the navigation system 104 (the vehicle speed can be received by other means, for example from measurements taken by the vehicle itself). Said first communication means 600 can include a CANbus communications port.

[0086] The device 100 can comprise second communication means 602 for receiving the ignition (IG) signal from the ignition system 102 or from energy management system itself of the vehicle 108. In a preferred embodiment, said second communication means 602 comprise a digital input port.

[0087] The device 100 comprises data processing means 604, for example a DSP unit or a microcontroller with high computational capacity for performing the different steps of the calculation. Said data processing means 604 have, or have access to, data storage means 606, for example a RAM memory and an EPROM memory.

[0088] The device has communication means for communicating with the energy management system of the vehicle 108, for example through a CANbus communications port 608. It can also have a digital output port 610.

Claims

1. A method for the on-line prediction of the driving cycle in an automotive vehicle, characterized in that it comprises:

- a step of data pre-processing (200), which in turn comprises:

• receiving the vehicle speed (V_sp);

• receiving traffic information (HTI) corresponding to the expected path for the vehicle within at least one prediction horizon (H) considered;

• obtaining (212) a reference driving cycle (V_pat) corresponding to the expected path within at least said prediction horizon (H) from the traffic information (HTI) received;

• calculating (208) the deviation (DV_sp) of the vehicle speed (V_sp) with respect to the reference driving cycle (V_pat);

- a step of data processing by means of a neural network (202), which comprises recursively obtaining the expected' deviations (D*V_sp) for the prediction horizon (H), using for that purpose the deviations of speed (NN_DV_sp) previously calculated and corresponding to the recent past in a delay sample number (DSN) as well as information relating to the reference driving cycle (NN_V_pat) containing information belonging both to the recent past in a delay sample number (DSN) and to the near future in a future sample number (FSN) as inputs of the neural network (202);

- a step of data post-processing (204) which comprises obtaining the estimated speed (V*_sp) for said prediction horizon (H) from the expected deviations (D*V_sp) and the reference driving cycle (V_pat) for the prediction horizon (H).

2. The method according to claim 1, characterized in that the step of data pre-processing (200) comprises receiving traffic events information (HTEI) corresponding to the expected path for the vehicle within at least the prediction horizon (H), and where said traffic events information (HTEI) received is also used for obtaining the reference driving cycle (V_pat).

3. The method according to any of the preceding claims, characterized in that the neural network (202) is a previously trained recurrent dynamic neural network with NARX topology.

4. The method according to any of the preceding claims, characterized in that the vehicle speed (V_sp) is sampled in the step of data pre-processing (200) according to a specific sampling time (ST); and where obtaining (212) the reference pattern cycle (V_pat) and the calculation of the deviation (DV_sp) of the vehicle speed (V_sp) with respect to the reference driving cycle (V_pat) are performed for each sampling time (ST).

5. The method according to any of the preceding claims, characterized in that the information relating to the reference driving cycle (V_pat) comprises a pattern speed moved forward a future sample number (FSN), which is equivalent to the vision distance of the driver and the anticipation of the driver with respect to future traffic situation changes.

6. The method according to any of the preceding claims, characterized in that the traffic information (HTI) additionally includes at least one of the following pieces of information:

- the speed limits;

- information of the type of thoroughfare;

- the road slopes;

- the traffic signals of the expected path.

7. The method according to any of the preceding claims, characterized in that the traffic events information (HTEI) includes information relating to at least one of the following:

- traffic state;

- speed limits due to road works;

- visibility conditions;

- road surface conditions.

8. The method according to any of the preceding claims, characterized in that the traffic information (HTI) and the traffic events information (HTEI) are received within the interval [p, p+H], p being the current vehicle position and H the selected prediction horizon.

9. The method according to any of the preceding claims, characterized in that it comprises obtaining the driving style (DS) of the driver of the vehicle according to calculations depending on a parameter relating to the driving style calculation mode (DSCM) selected, where the calculation modes are based on at least one of the following:

- calculation based on Fourier transform of a vector formed by the vehicle speed values corresponding to the recent past;

- calculation based on the mean speed variation over a period of time;

- calculation based on driver reaction times.

10. A device for the on-line prediction of the driving cycle in an automotive vehicle, characterized in that it comprises:

- communication means (600) configured for receiving the vehicle speed (V_sp) and for receiving traffic information (HTI) corresponding to the expected path for the vehicle within at least one prediction horizon (H) considered from a navigation system (104);

- data processing means (604) configured for:

• obtaining (212) a reference driving cycle (V_pat) corresponding to the expected path within at least said prediction horizon (H) from the traffic information (HTI) received by the communication means;

• calculating (208) the deviation (DV_sp) of the vehicle speed (V_sp) with respect to the reference driving cycle (V_pat);

• recursively obtaining the expected deviations (D*V_sp) for the prediction horizon (H) by means of a neural network (202), using for that purpose the deviations of speed (NN_DV_sp) previously calculated and corresponding to the recent past in a delay sample number (DSN) as well as information relating to the reference driving cycle (NN_V_pat) containing information belonging both to the recent past in a delay sample number (DSN) and to the near future in a future sample number (FSN) as inputs of the neural network (202);

• obtaining the estimated speed (V*_sp) for said prediction horizon (H) from the expected deviations (D*V_sp) and the reference driving cycle (V_pat) for the prediction horizon (H).

11. The device according to claim 10, characterized in that the communication means (600) are additionally configured for receiving traffic events information (HTEI) corresponding to the expected path for the vehicle within at least the prediction horizon (H) from the navigation system (104), and where the data processing means (604) are configured for obtaining the reference driving cycle (V_pat) also using said traffic events information (HTEI) received by the communication means.

12. The device according to any of claims 10 to 11, characterized in that the neural network (202) is a previously trained recurrent dynamic neural network with NARX topology.

13. The device according to any of claims 10 to 12, characterized in that the data processing means (604) are configured for sampling the vehicle speed (V_sp) according to a specific sampling time (ST) and for obtaining (212) the reference pattern cycle (V_pat) and the calculation of the deviation (DV_sp) of the vehicle speed (V_sp) with respect to the reference driving cycle (V_pat) for each sampling time (ST).

14. The device according to any of claims 10 to 13, characterized in that the information relating to the reference driving cycle (V_pat) comprises a pattern speed moved forward a future sample number (FSN), which is equivalent to the vision distance of the driver and the anticipation of the driver with respect to future traffic situation changes.

15. The device according to any of claims 10 to 14, characterized in that the data processing means are configured for performing the prediction calculation while the vehicle is circulating and every time the vehicle advances by a selected distance by means of a parameter (H_Resol).

Drawing