WO2024028343A1

WO2024028343A1 - Computer-implemented methods for forecasting availability of an infrastructure component and route planning based on the forecast

Info

Publication number: WO2024028343A1
Application number: PCT/EP2023/071322
Authority: WO
Inventors: Rong Su; Ruikang LUO; Yaofeng SONG; Liping Huang; Han Zhao
Original assignee: Continental Automotive Technologies GmbH; Nanyang Technological University
Current assignee: Aumovio Germany GmbH; Nanyang Technological University
Priority date: 2022-08-04
Filing date: 2023-08-01
Publication date: 2024-02-08
Anticipated expiration: 2025-02-04
Also published as: DE102022119607A1

Abstract

In order to improve infrastructure use or route planning, a forecasting system (10) is proposed. The forecasting system (10) is configured for forecasting an availability score (Y_i) of infrastructure components within a road network that is represented by road network structure data, the system comprising: a) augmenting means (12) that are configured for augmenting historical availability data (X_i) of the infrastructure component and road network structure data with at least one attribute data matrix (α, β) so as to obtain attribute augmented availability data (E_i), wherein the historical availability data (X_i) is indicative of an availability score of the infrastructure component at different points in time; b) a first machine learning model (14) that includes at least one graph convolutional network layer that is configured to determine spatial dependency output data that is indicative of the spatial dependency of the historical availability data (X_i) based on the attribute augmented availability data (E_i) and the road network structure data; c) a second machine learning model (16) that includes an informer layer that is trained to determine an availability score of the infrastructure component based on the spatial dependency output data, wherein the informer layer includes at least one encoder (20) and at least one decoder (22), wherein the encoder (20) determines a feature map (28) based on the spatial dependency output data, wherein the feature map (28) is fed into each decoder (22).

Description

DESCRIPTION

Computer-implemented methods for forecasting availability of an infrastructure component and route planning based on the forecast

TECHNICAL FIELD

The invention relates to a computer-implemented method for forecasting an availability of an infrastructure component, such as an electrical vehicle charging station.

BACKGROUND

Accurate traffic information forecasting plays an important role in the smart city management. Generally speaking, traffic information contains link speed, traffic flow, vehicle density, travelling time, facility usage condition and so on. As the rapid development of electric vehicle (EV) technologies continue, the proportion of EVs is increasing. However, limited endurance and charging stations, and much longer charging time compared with short refueling time for petrol cars is a source for mileage anxiety for EV drivers. As one of the most significant infrastructures of the EV infrastructure, EV charging stations attract more attention recently. Some studies show that EV charging behavior has periodicity, thereby an accurate EV charging station usage forecasting system can effectively alleviate range anxiety and improve road efficiency.

Another benefit from huge number of smart sensors, real-time station level monitoring has been realized. Most canonical facility usage condition prediction methods are dependent on historical traffic features to make prediction. However, EV charging station availability is generally much more complex than other time-series forecasting issues due to the future availability not only depending on past availability values, but also being related with topological relationships or comprehensive external influence. For example, within the campus or a central business district road section, the usage of a charging station can be highly affected by the commute time. An obvious rise of availability can be observed around offduty time, which is typically the opposite inside a residential area, even though the road structure may be similar. Another example is that bad weather, such as heavy rain, can reduce and delay people’s commute, and further affect charging station usage.

With the known methods it is quite a challenge to take into consideration the randomness caused by these external factors. As the development of deep learning technologies progresses, several forecasting methods has been proposed to solve this issue, such as the Auto regressive and Integrated Moving Average (ARIMA) method, the Convolutional Neural Network (CNN) method, the Long Short-Term Memory (LSTM) method, the Graph Convolution Network (GCN) method and Transformer based methods. Each algorithm has its own strengths and limitations. Most of the known models do not consider the external factors and correspondingly have a frailer perception of said external factors.

Additional reference is made to the following documents:

[1] R. Luo and R. Su, “Traffic signal transition time prediction based on aerial captures during peak hours,” in 2020 16th International Conference on Control, Automation, Robotics and Vision (ICARCV). IEEE, 2020, pp. 104-110.

[2] S. Storandt and S. Funke, “Cruising with a battery-powered vehicle and not getting stranded,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 26, no. 1 , 2012.

[3] K. Qian, C. Zhou, M. Allan, and Y. Yuan, “Load model for prediction of electric vehicle charging demand,” in 2010 International Conference on Power System Technology. IEEE, 2010, pp. 1-6.

[4] Y. Xiong, J. Gan, B. An, C. Miao, and A. L. Bazzan, “Optimal electric vehicle fast charging station placement based on game theoretical framework,” IEEE Transactions on Intelligent Transportation Systems, vol. 19, no. 8, pp. 2493-2504, 2017.

[5] G. Alface, J. C. Ferreira, and R. Pereira, “Electric vehicle charging process and parking guidance app,” Energies, vol. 12, no. 11 , p. 2123, 2019.

[6] J. Zhu, Q. Wang, C. Tao, H. Deng, L. Zhao, and H. Li, “Ast-gcn: Attribute- augmented spatiotemporal graph convolutional network for traffic forecasting,” IEEE Access, vol. 9, pp. 35 973-35 983, 2021. [7] C.-W. Huang, C.-T. Chiang, and Q. Li, “A study of deep learning networks on mobile traffic forecasting,” in 2017 IEEE 28th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC). IEEE, 2017, pp. 1-6.

[8] S. Siami-Namini, N. Tavakoli, and A. S. Namin, “A comparison of arima and Istm in forecasting time series,” in 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA). IEEE, 2018, pp. 1394-1401.

[9] W. Jiang and J. Luo, “Graph neural network for traffic forecasting: A survey,” arXiv preprint arXiv:2101 .11174, 2021 .

[10] J. Yang, H. Chen, Y. Xu, Z. Shi, R. Luo, L. Xie, and R. Su, “Domain adaptation for degraded remote scene classification,” in 2020 16th International Conference on Control, Automation, Robotics and Vision (ICARCV). IEEE, 2020, pp. 111-117.

[11] H. Zhao, H. Yang, Y. Wang, D. Wang, and R. Su, “Attention based graph bi-lstm networks for traffic forecasting,” in 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2020, pp. 1-6.

[12] X. Luo, D. Li, and S. Zhang, “Traffic flow prediction during the holidays based on dft and svr,” Journal of Sensors, vol. 2019, 2019.

[13] H. Zhao, H. Yang, Y. Wang, R. Su, and D. Wang, “Domain-adversarialbased temporal graph convolutional network for traffic flow prediction problem,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021 , pp. 1365-1370.

[14] H. Yang, Y. Wang, H. Zhao, J. Zhu, and D. Wang, “Real-time traffic incident detection using an autoencoder model,” in 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2020, pp. 1-6.

[15] D. Zhang and M. R. Kabuka, “Combining weather condition data to predict traffic flow: A gru-based deep learning approach,” IET Intelligent Transport Systems, vol.

12, no. 7, pp. 578-585, 2018.

[16] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in neural information processing systems, 2017, pp. 5998- 6008.

[17] R. Luo, Y. Zhang, Y. Zhou, H. Chen, L. Yang, J. Yang, and R. Su, “Deep learning approach for long-term prediction of electric vehicle (ev) charging station availability,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021 , pp. 3334-3339. [18] B. Liao, J. Zhang, C. Wu, D. Mcllwraith, T. Chen, S. Yang, Y. Guo, and F. Wu, “Deep sequence learning with auxiliary information for traffic prediction,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018, pp. 537-546.

[19] L. Zhao, Y. Song, C. Zhang, Y. Liu, P. Wang, T. Lin, M. Deng, and H. Li, “T-gcn: A temporal graph convolutional network for traffic prediction,” IEEE Transactions on Intelligent Transportation Systems, vol. 21 , no. 9, pp. 3848-3858, 2019.

[20] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv: 1609.02907, 2016.

[21] Y.-H. H. Tsai, S. Bai, M. Yamada, L.-P. Morency, and R. Salakhutdinov, “Transformer dissection: A unified understanding of transformer’s attention via the lens of kernel,” arXiv preprint arXiv: 1908.11775, 2019.

[22] H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proceedings of AAAI, 2021.

SUMMARY OF THE INVENTION

It is the object of the invention to improve infrastructure use by vehicles. The object is achieved by the subject-matter of the independent claims. Preferred embodiments are subject-matter of the dependent claims.

The invention provides a computer-implemented method for forecasting an availability score of infrastructure components within a road network that is represented by road network structure data, the method comprising: a) augmenting historical availability data of the infrastructure component and road network structure data with at least one attribute data matrix so as to obtain attribute augmented availability data, wherein the historical availability data is indicative of an availability score of the infrastructure component at different points in time; b) feeding the attribute augmented availability data and the road network structure data into a first machine learning model that includes at least one graph convolutional network layer that is configured to determine spatial dependency output data that is indicative of the spatial dependency of the historical availability data; c) feeding the spatial dependency output data into a second machine learning model that includes an informer layer that is trained to determine an availability score of the infrastructure component, wherein the informer layer includes at least one encoder and at least one decoder, wherein the encoder determines a feature map based on the spatial dependency output data, wherein the feature map is fed into each decoder.

Preferably, in step a) a first attribute data matrix represents static external factors that are associated with each infrastructure component, wherein the static external factors are determined by time-independent properties of the surroundings of the respective infrastructure component.

Preferably, in step a) a second attribute data matrix represents dynamic external factors that are associated with each infrastructure component and/or the road network, wherein the dynamic external factors are determined by time-dependent events that occur in the environment in which the respective infrastructure component is arranged.

Preferably, a time window of a predetermined length is chosen, and only events within that time window are augmented to the historical availability data.

Preferably, in step c) the informer layer includes a first encoder and a last encoder, wherein the output of step b) is fed to the first encoder and the output of the first encoder is fed to another encoder, wherein the feature map is output by the last encoder.

Preferably, in step c) the informer layer includes a first decoder and a last decoder, wherein the output of the last decoder is fed to a fully connected output layer.

Preferably, in step c) the spatial dependency output data are partially masked in a consecutive manner and the partially consecutively masked data are fed into the decoder or the first decoder. Preferably, in step c) each encoder includes a first self-attention pyramid and a second self-attention pyramid, wherein the second self-attention pyramid has fewer inputs than the first self-attention pyramid, and each self-attention pyramid generates a semi-feature map that are concatenated into the output feature map.

Preferably, the second self-attention pyramid has half the inputs of the first selfattention pyramid.

The invention provides a computer-implemented method for planning a route of a vehicle, the method comprising: a) determining an initial route from a starting position to a destination; b) determining a route length of the initial route and comparing the route length with a remaining distance the vehicle is able to travel due to fuel consumption; c) if the route length is determined to be longer than the remaining distance, performing a previously described method to obtain at least one availability score, determining an intermediate stop based on the availability score, and modifying the initial route to include the intermediate stop to obtain a final route; d) generating a control signal that causes the vehicle to inform a driver of the final route or that causes the vehicle to follow the final route.

The invention provides a computer program that includes instructions that, when executed by a computer, cause the computer to perform one, some, or all steps of a previously described method.

The invention provides a computer readable storage medium or a data carrier signal that includes the computer program.

The invention provides a forecasting system configured for forecasting an availability score of infrastructure components within a road network that is represented by road network structure data, the system comprising: a) augmenting means that are configured for augmenting historical availability data of the infrastructure component and road network structure data with at least one attribute data matrix so as to obtain attribute augmented availability data, wherein the historical availability data is indicative of an availability score of the infrastructure component at different points in time; b) a first machine learning model that includes at least one graph convolutional network layer that is configured to determine spatial dependency output data that is indicative of the spatial dependency of the historical availability data based on the attribute augmented availability data and the road network structure data; c) a second machine learning model that includes an informer layer that is trained to determine an availability score of the infrastructure component based on the spatial dependency output data, wherein the informer layer includes at least one encoder and at least one decoder, wherein the encoder determines a feature map based on the spatial dependency output data, wherein the feature map is fed into each decoder.

The invention provides a vehicle route planning device for a vehicle, the device comprising a forecasting system.

One idea is to build a neural network that is capable of extracting both spatial- temporal information and external influence in order to predict the charging station usage condition, i.e. , availability. An Attribute-Augmented Spatial-Temporal Graph Informer Network (AST-GIN) is used as a machine learning model in determining the availability score. The AST-GIN includes an Attribute Augmentation Unit (A2U), a Graph Convolutional Network (GCN) and an Informer Network. The disclosed AST- GIN model can well consider the dynamic and static external factor influence on the EV charging station in contrast to known models. As will be described later, the AST- GIN is compared with known baselines by testing with a real-world dataset collected at Dundee.

For traffic forecasting, the approaches have undergone several stages and methods could be generally divided into two types: canonical models and deep learning-based models. Canonical forecasting models usually build mathematical models and treat traffic behavior as a conditional process. There are many famous models, such as the Historical Average (HA) model, K-nearest Neighbor model, ARIMA model and Support Vector Regression (SVR) model. Most of these models consider the trend of data and make the strong assumption that time-series data is stable, which makes these models difficult to respond to a rapid change of inputs. Recently, deep learning-based forecasting methods have been widely applied to predict time-series. These models are usually capable to extract nonlinear relationships from the input sequence and include a Recurrent Neural Network (RNN) model, Stacked Autoencoding Neural Network (SAE), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) model, Transformer and their variants. The modesl are more efficient to extract the temporal information than canonical forecasting models.

As mentioned above, external factors influence future availability of EV stations. Existing methods can be further improved by considering external information fluence. This is done with the AST-GIN for EV charging station availability forecasting, which integrates both spatial-temporal and external factors as input to enhance the model’s perception capability during predicting.

The goal of the invention is to predict a future availability of an EV charging statinon based on historical states and associated information. Based on the a priori knowledge introduced, the demand of EVs has a strong periodicity and external factors have a high correlation with the usage of EVs.

Finding an availability function that estimates an availability score for a charging station is highly potential. As an example, the historical aggregated EV station availability data and two external factors, weather condition and POI information are used to exemplify the idea.

The road network that connects different charging stations is treated as a weighted undirected graph G = {V, E}. The EV stations are represented by the nodes V inside the graph. E represents the graph edge set representing connectivity between the stations. The corresponding adjacency matrix A can be constructed based on node and edge information. With a road map, based on the latitude and longitude of the charging stations, the road distance between EV stations can be estimated. The adjacency matrix elements are preferably calculated using a Gaussian kernel weighting function having a predetermined cut-off. In other words only if the distance between two stations is smaller than a predetermined cut-off distance, then the matrix element is greater than 0.

Another element is a traffic feature matrix Xi that contains high-dimension information of EV station availability. In other words the traffic feature matrix includes the historical aggregate availability score of each charging station that is modeled.

Thus, at time i, the known L time steps [X-L, X-L+I , ... , Xi] historical states are used as partial inputs to predict the next M steps states [Yj+i , ... ,Yj+M].

Further, the influence of external factors is regarded as affiliated attributes matrix F. These factors construct an attribute matrix [Fi , F2, ... , Fi], where I is the dimension of attribute information. At time i, the set of the j-th affiliated information is Fj = [ji-L,ji-L+i ,

Ji]

The invention solves the technical problem of EV charging station availability forecasting by considering external factors and refining the relationship function f based on the historical usage data X, attribute matrix F and road graph structure G, to obtain the future usage values Y.

The AST-GIN model contains an Attribute Augmentation Unit (A2Unit), which can integrate the external information, a GCN layer and an Informer layer.

The historical time-series data and external data are fed into the A2Unit for attribute augmentation. Then, the processed information is fed into the GCN layer for spatial information extraction. Finally, the Informer layer will take outputs from the GCN layer to extract the temporal dependencies.

To comprehensively take external factors influence into consideration, dynamic attributes and static attributes are selected respectively for the objective region. EV stations historical availability matrix X, road structure G and two types of attribute matrix are integrated into the A2Unit for augmentation. A static attribute matrix a contains p categories of attributes which are time-invariant. This may refer to the static environment around the charging station, e.g. whether there is a supermarket or other services that one can use while the car is charging.

Similarly [3 represents w different dynamic attributes, which are changing over time. This includes dynamic environmental phenomena such as weather, traffic situation, etc. To aggregate the cumulative influence of dynamic attributes, a historical window of length L is selected. Thus, the final augmented matrix processed by A2Unit at time i is stated as Ei = [Xi, a, PM, PM+I , ... Pi], and the same processing procedure is applied for every time stamp inside traffic feature matrix X. In other words, for each time i, the augmented matrix Ei is obtained by appending to the matrix Xi additional columns that include the static attributes a and a historical time window of the dynamical attributes p.

The GCN layer is used to extract the spatial dependencies of the input data. The convolution is given by Hi+i = O(D’^1/2AD’^{1 /2}HIWI), where o is the activation function, D is the diagonal degree matrix, A is the adjacency matrix with self-loops, Wi is the weight matrix of convolutional layer I, Hi+i is the output of the convolutional layer, Hi is the input, which is for Ho the augmented matrix E.

The informer layer is used to obtain global temporal dependency while forecasting. The informer layer employs an encoder-decoder architecture. The model improves the efficiency of the query’s attention by measuring the sparsity and decreasing the dimension of query value in the Multi-head ProbSparse Self-attention block.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the invention are described in more detail with reference to the accompanying schematic drawings.

Fig. 1 depicts an embodiment of an AST-GIN model;

Fig. 2 depicts an embodiment of an informer layer;

Fig. 3 depicts a table of experimental results; and

Fig. 4 depicts a diagram illustrating accuracy of the compared models. DETAILED DESCRIPTION OF EMBODIMENT

Referring to Fig. 1 , a forecasting system 10 is partially depicted. The forecasting system 10 is configured for forecasting an availability score of an infrastructure component, such as an electric charging station for electric vehicles. Usually a plurality of charging stations are distributed in an area and connected by a road network. It should be noted that the invention is described based on electric charging infrastructure, but the ideas discussed may be readily applied to other infrastructure components.

Each charging station has is installed in an environment that may have static and/or dynamic attributes. Static attributes are time-independent, e.g., surrounding shops or services, whereas dynamic attributes change over time, such as weather.

The forecasting system 10 is fed with historical availability data Xi for each charging station. The historical availability data may be indicative of which amount of individual chargers are available at the charging station.

The forecasting system 10 is fed a static attribute data matrix a. The static attribute data matrix a is indicative of the static environment around the electric charger.

The forecasting system 10 is fed with a dynamic attribute data matrix [3 that includes entries for each point in time i back into the past for a time window L, which can be chosen. For example, the dynamic attribute matrix [3 includes weather data for each charging station for the last L hours for example.

The forecasting system 10 includes an augmentation means 12. The augmentation means 12 receive the historical availability data Xi and the attribute data matrices a, [3. The augmentation means 12 concatenate these input data into attribute augmented availability data Ei.

The forecasting system 10 includes a first machine learning model 14 having at least one graph convolutional network layer (GCN). The road network that connects the charging stations is represented as an undirected graph G. Each charging station is represented by a node V and the road network is represented by the edges E. The road network data is represented by an adjacency matrix A that is generated from the locations of each charging station and the connecting roads. The adjacency matrix A is using a Gaussian kernel function with a predetermined cut-off. The GCN extracts the spatial dependency of the input data and feeds the result to a second machine learning model 16.

Referring to Fig. 2, the forecasting system 10 comprises the second machine learning model 16 that includes at least one informer layer 18. The informer layer 18 has at least one encoder 20 and at least one decoder 22.

The encoder 20 comprises a first self-attention pyramid 24 and a second selfattention pyramid 26 that has half the inputs of the first self-attention pyramid 24.

Both the self-attention pyramids 24, 26 are configured as ProbSparse self-attention pyramids. The output of the self-attention pyramids 24, 26 is concatenated into a concatenated feature map 28.

The concatenated feature map 28 is fed to the decoder 22 into a multi-head attention block 30. The multi-head attention block 30 receives the output of a masked selfattention pyramid 32 that is fed with masked availability data 34. The decoder 22 has an output layer that normalizes the decoder output so as to output an availability score for each station charger.

The final output of the forecasting system 10 is an expected availability score timeseries that includes for each point i within a predetermined future time frame, e.g., from 30 minutes to 120 minutes in 30 minute increments a prediction of the availability score Y, of each charging station in the road network.

This output may be fed to a route planning system that is known and capable of planning a route based on remaining distance in a battery charge. In addition, the route is also determined based on the availability score Y of the charging stations in the vicinity of the route. The route planning system may determine an initial route based on user input and/or position data. Then the charge that is needed for driving the route is determined. If the remaining distance in the energy storage of the vehicle is insufficient to reach the destination, the route is modified to include at least one charging station based on the predicted availability score Y, at the time of arrival at the respective charging station. As a result, the electric infrastructure and available ressources can be used more efficiently. In addition, mileage anxiety of the driver can be reduced.

To evaluate the AST-GIN model performance, necessary experiments have been done on the EV charging station availability dataset. We choose 5 efficient timeseries forecasting baseline models for comparison. During the experiment, the performance of AST-GIN model with static external factor only, the model with dynamic external factor only and the model with both static and dynamic factors would be evaluated separately.

Dundee EV charging dataset: this dataset (publicly available at the website) is a record of the EV charging behaviors in Dundee, Scotland. There are 57 charging points in Dundee which could be divided into 3 types including slow chargers, fast chargers and rapid chargers. Totally 3 valid datasets recorded in 3 different time periods including 01/09/17 to 01/12/17, 02/12/17 to 02/03/18 and 05/03/18 to 05/06/18 are accessible. Meanwhile, the geographical location of all the charging points is also provided.

In the experiment, the dataset we used is recorded during 05/03/18 to 05/06/18. There are 16773 charging sessions recorded in total and each of the charging session records contains charging point ID, charging connector ID, starting and ending charging time, total consumed electric power, geographical location, and the type of the charging point. There are 40 slow chargers with 5894 charging sessions recorded, 8 fast chargers with 1416 charging sessions recorded and 9 rapid chargers with 9463 charging sessions recorded.

Moreover, the Dundee weather dataset, which is available at the website, is a weather record in Dundee city. The weather is recorded every hour and each record includes the general description of the weather, temperature, wind, humidity and barometer.

The availability of a specific charging station at the p-th session is calculated every y j

30 minutes and it can be described as a_p = 1 - — — — - — where Z t_p is the total

charging duration within a specific session and Mconnector is the number of charging connectors at that station.

In the experiment, the surroundings of the charging points are classified into 8 types including transportation services, catering services, shopping services, education services, accommodation, medical services, living services and other. The category of surroundings with the largest proportion would be labeled as the POI value of a charging point based on the geographical location of it.

The weather of Dundee is divided into 5 types which includes sunny, cloudy, foggy, light rainy, and heavy rainy with different labels from 1 to 5. Since the time interval of the weather data from the source dataset is 1 hour, the weather in the covered period is regarded the same which means that if the weather at 17:50 is recorded as sunny, the weather at 18:20 would be labeled as sunny.

In the experiment a 3-layer GCN structure is used. For each Informer block, the number of encoder layers is 2 while the number of decoder layers is 3. During the training, 50% data is randomly divided for training, 33% data is selected for evaluation and the rest 17% data is utilized for testing purpose. The network is optimized using Adam optimizer. The learning rate starts from e^-4 and decays 10 times smaller every 2 epochs. The total number of epochs is 50 with an early stop determination. The batch size is chosen as 32 and the loss function is mean square error. The whole network is trained on a GPU RTX3060

In the experiment commonly used metrics are selected to evaluate models forecasting performance, namely Root Mean Squared Error (RMSE), Coefficient of Determination (R²), Explained Variance Score (var), Mean Absolute Error (MAE), and Accuracy. The inventors used 5 state of art baselines to compare the performance with the AST-GIN model according to the invention. The baselines include GRU, LSTM, Transformer, Informer and Spatial Temporal Transformer Network (STTN). Based on the 30-minute time interval of the EV charging availability dataset, the selected models are deployed to predict the availability in the next 30-min, 60-min, 90-min and 120-min horizons. The numerical results are shown in Fig. 3.

In the short-term EV charging station availability forecasting, horizons of 30 minutes for example, the AST-GIN model effectively captures the temporal dependance of the data and outperforms other baseline models with a higher accuracy. In the 90 minutes forecasting, the AST-GIN model with the dynamic external factors achieves an accuracy of 0.8388, while the best model, STTN, among the baseline models achieve an accuracy of 0.7520. AST-GIN outperforms STTN in the 90 minutes horizon with a 11 .54% higher accuracy. In the longterm forecasting, 120 minutes for example, the AST-GIN still has the best performance with the highest accuracy compared to the other baseline models. The accuracy of AST-GIN, 0.7507, is 8.97% higher than the best model, STTN, among baseline models. The consolidated result is shown in Fig. 4.

Among the 3 kinds of used external factors which includes static factors only, dynamic factors only and the combination of both of the factors, the use of external factors combination leads to a better performance in general.

In this disclosure, the deep learning model AST-GIN is described and verified for EV charging station availability forecasting considering the influence of external factors. The model contains an A2Unit layer, at least one GCN layer and at least one Informer layer to augment time-series traffic features and extract spatial-temporal dependencies of an EV charging station usage condition. AST-GIN and baselines are tested on the data collected in Dundee City. Experiments show that AST-GIN has the better forecasting capability over various horizons and metrics. To summarize, AST- GIN can effectively consider comprehensive external attribute influence and predict EV charging station usage condition. AST-GIN can be applied to the applications of EV charging systems, to accurately predict charging demand and supply status to suggest drivers a better charging behavior arrangement, or further provide key information towards a scheduling system or route planning system .

REFERENCE SIGNS

10 forecasting system

12 augmentation means

14 first machine learning model

16 second machine learning model

18 informer layer

20 encoder

22 decoder

24 first self-attention pyramid

26 second self-attention pyramid

28 concatenated feature map

30 multi-head attention block

32 masked self-attention pyramid

34 masked availability data

A adjacency matrix

E edges

Ei attribute augmented availability data

G undirected graph

V node

Xi historical availability data a static attribute data matrix

[3 dynamic attribute data matrix

Claims

1 . A computer-implemented method for forecasting an availability score of infrastructure components within a road network that is represented by road network structure data, the method comprising: a) augmenting historical availability data of the infrastructure component and road network structure data with at least one attribute data matrix so as to obtain attribute augmented availability data, wherein the historical availability data is indicative of an availability score of the infrastructure component at different points in time; b) feeding the attribute augmented availability data and the road network structure data into a first machine learning model that includes at least one graph convolutional network layer that is configured to determine spatial dependency output data that is indicative of the spatial dependency of the historical availability data; c) feeding the spatial dependency output data into a second machine learning model that includes an informer layer that is trained to determine an availability score of the infrastructure component, wherein the informer layer includes at least one encoder and at least one decoder, wherein the encoder determines a feature map based on the spatial dependency output data, wherein the feature map is fed into each decoder.

2. The method according to claim 1 , ch a racte ri zed i n th at in step a) a first attribute data matrix represents static external factors that are associated with each infrastructure component, wherein the static external factors are determined by timeindependent properties of the surroundings of the respective infrastructure component.

3. The method according to any of the preceding claims, ch a racte ri zed i n th at in step a) a second attribute data matrix represents dynamic external factors that are associated with each infrastructure component and/or the road network, wherein the dynamic external factors are determined by time-dependent events that occur in the environment in which the respective infrastructure component is arranged.

4. The method according to claim 3, ch a racte ri zed i n th at a time window of a predetermined length is chosen, and only events within that time window are augmented to the historical availability data.

5. The method according to any of the preceding claims, ch a racte ri zed i n th at in step c) the informer layer includes a first encoder and a last encoder, wherein the output of step b) is fed to the first encoder and the output of the first encoder is fed to another encoder, wherein the feature map is output by the last encoder.

6. The method according to any of the preceding claims, ch a racte ri zed i n th at in step c) the informer layer includes a first decoder and a last decoder, wherein the output of the last decoder is fed to a fully connected output layer.

7. The method according to any of the preceding claims, ch a racte ri zed i n th at in step c) the spatial dependency output data are partially masked in a consecutive manner and the partially consecutively masked data are fed into the decoder or the first decoder.

8. The method according to any of the preceding claims, ch a racte ri zed i n th at in step c) each encoder includes a first self-attention pyramid and a second selfattention pyramid, wherein the second self-attention pyramid has fewer inputs than the first self-attention pyramid, and each self-attention pyramid generates a semifeature map that are concatenated into the output feature map.

9. The method according to claim 8, ch a racte ri zed i n th at the second selfattention pyramid has half the inputs of the first self-attention pyramid.

10. A computer-implemented method for planning a route of a vehicle, the method comprising: a) determining an initial route from a starting position to a destination; b) determining a route length of the initial route and comparing the route length with a remaining distance the vehicle is able to travel due to fuel consumption; c) if the route length is determined to be longer than the remaining distance, performing a method according to any of the preceding claims to obtain at least one availability score, determining an intermediate stop based on the availability score, and modifying the initial route to include the intermediate stop to obtain a final route; d) generating a control signal that causes the vehicle to inform a driver of the final route or that causes the vehicle to follow the final route.

11 . A computer program that includes instructions that, when executed by a computer, cause the computer to perform one, some, or all steps of a method according to any of the preceding claims.

12. A computer readable storage medium or a data carrier signal that includes the computer program according to claim 11 .

13. A forecasting system configured for forecasting an availability score of infrastructure components within a road network that is represented by road network structure data, the system comprising: a) augmenting means that are configured for augmenting historical availability data of the infrastructure component and road network structure data with at least one attribute data matrix so as to obtain attribute augmented availability data, wherein the historical availability data is indicative of an availability score of the infrastructure component at different points in time; b) a first machine learning model that includes at least one graph convolutional network layer that is configured to determine spatial dependency output data that is indicative of the spatial dependency of the historical availability data based on the attribute augmented availability data and the road network structure data; c) a second machine learning model that includes an informer layer that is trained to determine an availability score of the infrastructure component based on the spatial dependency output data, wherein the informer layer includes at least one encoder and at least one decoder, wherein the encoder determines a feature map based on the spatial dependency output data, wherein the feature map is fed into each decoder.

14. A vehicle route planning device for a vehicle, the device comprising a forecasting system according to claim 13.