Spatio-temporal graph convolution-based electric vehicle charging demand prediction method, and device therefor

Abstract

Disclosed are a spatio-temporal graph convolution-based electric vehicle charging demand prediction method, and a device therefor. The spatio-temporal graph convolution-based electric vehicle charging demand prediction method comprises the steps of: (a) dividing data related to an electric vehicle charging station (EVCS) into a dynamic graph and time series data; (b) generating spatial feature information by applying the dynamic graph and the time series data to a relational graph convolution-based first deep learning model; (c) generating temporal feature information by applying the dynamic graph and the time series data to a bidirectional LSTM-based second deep learning model; and (d) outputting a charging demand prediction result per time for the EVCS by combining the spatial feature information and the temporal feature information.

Description

Time-space graph convolution-based electric vehicle charging demand forecasting method and apparatus

[0001] The present invention relates to a time-space graph convolution-based electric vehicle charging demand forecasting method and apparatus.

[0002]

[0003] Electric vehicles (EVs) are garnering attention as a promising solution for reducing global CO2 emissions and mitigating climate change. According to the International Energy Agency (IEA), the number of EVs worldwide is expected to exceed 240 million by 2030. While the rapid popularity of EVs has led to increased demand for charging, uncertainty regarding the demand for electric vehicle charging stations (EVCS) arises because EV users' charging behavior possesses probabilistic characteristics in terms of time and space. This poses risks to charging variability and operations aimed at maximizing profits. As retail power markets become more liberalized and customers gain access to various options for managing electricity costs, forecasting future EVCS charging demand is essential for participating in the market and efficiently accommodating charging needs. Therefore, developing accurate forecasting models is critical to address uncertainties related to EV demand, ensure reliable and economical operations, and further expand the adoption of EVCS.

[0004] Various prediction models developed to forecast the charging demand of electric vehicle charging stations (EVCS) include statistical methods, machine learning-based methods, and deep learning-based methods. Statistical models primarily include Monte Carlo simulation, fuzzy linear regression, quantile linear regression, autoregressive integral moving average (ARIMA), and Kalman filter methods. However, these statistical models assume that the original data is stationary, so they have limitations in handling nonlinear problems. Furthermore, they ignore potential dependencies between multiple time series under complex charging conditions. Although a method combining fuzzy sets and model predictive control has been proposed as an advanced statistical approach, obtaining fuzzy membership functions can be difficult, and it may not be suitable for handling highly uncertain situations involving many variables.

[0005] To overcome the limitations of statistical models, machine learning-based models have been developed to better capture complex non-linear relationships. These machine learning models may include support vector regression, kernel-based methods, nearest neighbor analysis, and random forests. While these models are more efficient than statistical models, their ability to map attributes is limited. This limitation is particularly pronounced when dealing with high-dimensional EV time series data, where the sample size is large or the relationships between attributes are more complex. Even as the number of data samples or attributes increases, these models generally remain stable and show only slight changes in accuracy.

[0006] Deep learning-based models demonstrate superior performance because they can learn from large amounts of complex data. Deep learning-based models, such as Recurrent Neural Networks (RNNs) and their variations like Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are widely used for forecasting EV charging demand. However, RNNs and their variations only consider the temporal dependency of charging demand while ignoring spatial dependency within EVCS networks; furthermore, sparse relationship patterns exist between different nodes, and spatial dependency significantly impacts demand forecasting. For example, EVCSs within a specific block may exhibit similar consumption patterns because they are subject to similar traffic patterns and weather influences.

[0007] To simultaneously capture temporal and spatial dependencies, Convolutional Neural Networks (CNNs) and their variations were utilized alongside RNN-based models. Specifically, CNNs combined with Bi-Gated Recurrent Units (Bi-GRUs) were used to capture spatial-temporal dependencies for demand forecasting. Additionally, Fully Convolutional Neural Networks (FCN-CE-LSTM), combining Extended Conv-LSTMs and Residual Networks, were proposed to capture spatial-temporal dependencies for demand forecasting. However, CNN-based models can only consider the absolute spatial relationships between EVCSs in a two-dimensional Euclidean space. In particular, CNNs require regular grid data as input, which is unsuitable because typical EVCS points are irregularly distributed on a map. In this regard, to fit EVCS data to the CNN, it must be forcibly sampled into a regular grid format, which presents the problem of compromising the original spatial information of the EVCS. Therefore, the ideal method to represent EVCS data is to model it in a graph form that reflects its original spatial shape.

[0008] Accordingly, a new type of neural network called the Graph Convolutional Neural Network (GCN) was proposed. To map irregular data using a graph, the input data of the GCN consists of nodes and edges representing EVCS charging points and the relationships between them, respectively. Considering the unique graph structure of the EVCS network, the GCN maintains a realistic topology and can extract spatial dependencies between EVCS by aggregating node information through graph convolution. A GCN-LSTM model was proposed to predict traffic conditions by learning the interactions between roads within a traffic network. A gate space-time graph neural network was proposed for multi-bus load prediction. A pseudo-weighted space-time graph was constructed to reflect the coupling relationships between the loads of each bus.

[0009] However, conventional GCNs construct graphs using fixed adjacency matrices calculated based on prior knowledge. This approach can capture spatial-temporal dependencies using global fixed parameters, potentially overlooking heterogeneity in temporal and spatial dimensions. To overcome these issues, a self-adaptive adjacency matrix was proposed to capture hidden spatial dependencies between different nodes without prior knowledge, and a GCN-based RNN optimized to learn relationships between road zones from traffic data was proposed. Additionally, several GCN-RNNs (MGCRNNs) were used to capture spatial and other heterogeneous inter-regional correlations. A bidirectional adjacency matrix was developed to extract location-specific spatial interactions, and a heterogeneous spatial-temporal graph convolutional network was proposed to predict EV charging demand. Here, geographic and demand graphs were used to construct spatial correlations between different charging regions.

[0010] Meanwhile, various open-source forecasting tools such as the Global Change analysis model, the total power load model, and EVI-Pro Lite have recently been developed to forecast spatial and temporal traffic charging demand profiles. In particular, a spatially distributed statistical downscaling method has been proposed to forecast annual traffic energy as an hourly charging demand load profile. However, although these tools consider the spatial and temporal dependencies of EV charging demand, they are designed primarily for county, state, and balance authority scales, so there is a problem in that they cannot be directly used by EVCS aggregators to forecast individual EVCS charging demand and participate in the retail power market.

[0011] Despite recent advancements, existing models have four major limitations. First, while improvements to adjacency matrices have clarified data processing, their inter-node relationships remain static, ignoring other semantic elements that could be used to measure spatial relationships between nodes. In particular, EVCS demand is influenced not only by static attributes such as distance between nodes but also by dynamic attributes such as inter-station correlations, charging patterns, weather, day of the week, and holidays. These factors must be comprehensively considered alongside static attributes when constructing graphs. Furthermore, the factors influencing charging prediction are not clearly defined in existing adjacency matrices. Second, existing RNNs achieve maximum predictability by learning historical charging demand to capture temporal dependencies. However, EVCS demand also depends on external factors such as time of day, day of the week, historical EVCS demand conditions, and average charging rates. Third, to capture spatial and temporal dependencies, previous studies have sequentially combined GCN / CNN and RNN models. However, this sequential structure can distort dependencies during the transformation process of convolutional results, leading to information loss and increased prediction uncertainty. Therefore, a new deep learning structure is required to forecast EVCS demand that maintains the validity of spatial and temporal characteristics and integrates them complementarily. Finally, existing studies have not considered participation in the actual energy market to verify the practical applicability and robustness of the proposed model.

[0012]

[0013] The present invention is intended to provide a time-space graph convolution-based electric vehicle charging demand forecasting method and apparatus.

[0014] Furthermore, the present invention aims to provide a time-space graph convolution-based electric vehicle charging demand forecasting method and apparatus capable of simultaneously capturing temporal and spatial dependencies through a mutual residual graph convolution combined bidirectional model, which is an innovative parallel-structured charging demand forecasting model that enables EVCS owners to actively participate in the energy market.

[0015]

[0016] According to one aspect of the present invention, a time-space graph convolution-based electric vehicle charging demand forecasting method is provided.

[0017] According to one embodiment of the present invention, a time-space graph convolution-based electric vehicle charging demand forecasting method may be provided, comprising: (a) dividing EVCS (Electric vehicle charging station) related data into a dynamic graph and time series data; (b) applying the dynamic graph and the time series data to a first deep learning model based on relational graph convolution to generate spatial feature information; (c) applying the dynamic graph and the time series data to a second deep learning model based on bidirectional LSTM to generate temporal feature information; and (d) combining the spatial feature information and the temporal feature information to output an hourly charging demand forecasting result for the EVCS.

[0018] The dynamic graph above is a time series graph containing spatiotemporal correlations for each EVCS (Electric vehicle charging station) location and charging demand data over a continuous time series.

[0019] The above time series data may be an adjacency matrix generated using static attributes based on the EVCS location and dynamic attributes that affect the charging demand of the EVCS.

[0020] It can be generated by calculating the distance between EVCSs using the above-mentioned EVCS locations, constructing a distance matrix using the distance between EVCSs, constructing a dynamic feature correlation matrix based on the charging demand correlation with adjacent EVCSs based on Pearson correlation coefficients, constructing a charging convenience matrix based on weather information, and constructing a charging external factor matrix based on day of the week information regarding holidays and commuting, and then combining the distance matrix, the dynamic feature correlation matrix, the charging convenience matrix, and the charging external factor matrix.

[0021] The first deep learning model can derive residual elements using the time series graph and the mutual adjacency matrix, and generate spatial feature information by adding the residual elements to the result of applying the time series graph and the mutual adjacency matrix to a relational graph convolution-based model.

[0022] The first deep learning model comprises: a first graph convolution network block that receives the time series graph and the mutual adjacency matrix as input and extracts local spatial feature values ​​for local spatial dependencies that reflect the charging demand patterns of EVCSs located at adjacent distances; and a second graph convolution network block that extracts global spatial feature values ​​for global spatial dependencies that reflect the status and charging demand patterns of EVCSs located at distant distances using the local spatial feature values ​​which are the output of the first graph convolution network block, wherein spatial feature information can be generated by adding the residual element to the global spatial feature values.

[0023] The above time series data is a sequence of day-type trend features and may include time of day, day of week, past charging demand status and average charging rate.

[0024] The second deep learning model described above includes a multilayer feedforward neural network model that receives the time series data and generates an output vector reflecting the EVCS demand pattern; and a bidirectional LSTM model that receives the time series graph and outputs a temporal feature vector, wherein temporal feature information can be generated by combining the output vector of the multilayer feedforward neural network model and the temporal feature vector of the bidirectional LSTM model.

[0025]

[0026] According to another aspect of the present invention, an apparatus for performing a time-space graph convolution-based electric vehicle charging demand forecasting method is provided.

[0027] According to one embodiment of the present invention, a computing device may be provided comprising: a preprocessing unit that divides EVCS (Electric vehicle charging station) related data into a dynamic graph and time series data; and a prediction model unit that generates spatial feature information by applying the dynamic graph and the time series data to a first deep learning model based on relational graph convolution, generates temporal feature information by applying the dynamic graph and the time series data to a second deep learning model based on bidirectional LSTM, and then outputs an hourly charging demand prediction result for the EVCS by combining the spatial feature information and the temporal feature information.

[0028]

[0029] By providing a time-space graph convolution-based electric vehicle charging demand forecasting method and apparatus according to an embodiment of the present invention, temporal and spatial dependencies can be simultaneously captured through a mutual residual graph convolution combined bidirectional model, which is an innovative parallel-structured charging demand forecasting model that enables EVCS owners to actively participate in the energy market.

[0030]

[0031] FIG. 1 is a flowchart illustrating a time-space graph convolution-based electric vehicle charging demand forecasting method according to an embodiment of the present invention.

[0032] FIG. 2 is a drawing illustrating graph data according to an embodiment of the present invention.

[0033] FIG. 3 is a diagram illustrating the generation of mutually adjacent matrices according to an embodiment of the present invention.

[0034] FIG. 4 is a diagram illustrating the structure of a first deep learning model according to an embodiment of the present invention.

[0035] FIG. 5 is a diagram illustrating a bidirectional LSTM model according to an embodiment of the present invention.

[0036] FIG. 6 is a diagram illustrating the structure of a second deep learning model according to an embodiment of the present invention.

[0037] FIG. 7 is a diagram illustrating the overall structure of a charging demand prediction model according to one embodiment of the present invention.

[0038] FIG. 8 is a diagram illustrating a market operation framework according to an embodiment of the present invention.

[0039] FIG. 9 is a block diagram schematically illustrating the internal configuration of a computing device according to one embodiment of the present invention.

[0040]

[0041] As used in this specification, singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "composed" or "comprising" should not be interpreted as necessarily including all of the various components or steps described in the specification, and should be interpreted as meaning that some of the components or steps may be excluded, or that additional components or steps may be included. Furthermore, terms such as "...part," "module," etc., as used in the specification refer to a unit that processes at least one function or operation, which may be implemented in hardware or software, or a combination of hardware and software.

[0042] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0043]

[0044] FIG. 1 is a flowchart illustrating a time-space graph convolution-based electric vehicle charging demand forecasting method according to an embodiment of the present invention, FIG. 2 is a diagram illustrating graph data according to an embodiment of the present invention, FIG. 3 is a diagram illustrating mutual adjacency matrix generation according to an embodiment of the present invention, FIG. 4 is a diagram illustrating the structure of a first deep learning model according to an embodiment of the present invention, FIG. 5 is a diagram illustrating a bidirectional LSTM model according to an embodiment of the present invention, FIG. 6 is a diagram illustrating the structure of a second deep learning model according to an embodiment of the present invention, FIG. 7 is a diagram illustrating the overall structure of a charging demand forecasting model according to an embodiment of the present invention, and FIG. 8 is a diagram illustrating a market operation framework according to an embodiment of the present invention.

[0045] In step 110, the computing device (100) organizes EVCS-related data by separating it into dynamic graphs and time series data.

[0046] I will explain this in more detail.

[0047] An undirected graph is generated as G=(V, E, A) using EVCS location and charging demand data. For example, referring to Fig. 2(a), each EVCS is indicated by a green dot, and each green dot is considered an EVCS node; their relationship can be represented by G. Here, V={v1, v2, ↳, v m} represents the entire set of m EVCVS nodes, E represents the edge set, and A represents the adjacency matrix. Additionally, n represents the number of features for each node representing the spatial relationships between EVCS nodes.

[0048] In this way, after generating an undirected graph for each EVCS, the spatiotemporal correlations of the EVCS nodes in a continuous time series from time t to (t+1) can be represented as shown in (b) of Fig. 2. For example, v3 can obtain spatial correlations from nodes v1, v4, v2, and v6 that can potentially interact.

[0049] In Figure 2(b), the blue line represents the effect that node v3 has on itself during the next time step, which can be expressed as a time correlation.

[0050] Based on the graph transformation of EVCS demand data, the EVCS demand forecasting problem can be transformed into a forecast X(l+Δt) by considering the EVCS graph {X(t)| t=0, 1, Δt, l}. Here, l and t represent the last recording time and the forecast time interval, respectively.

[0051] The conventional adjacency matrix A is intuitive but incomplete, leading to biased prediction results. Furthermore, A has limitations in mining node association features as it ignores distance factors and correlations between EVCS nodes.

[0052] Furthermore, there is a disadvantage in that external factors that may affect EV user convenience and EVCS demand are not considered in accurate forecasting. For example, the correlation between two EVCS nodes does not increase even if the distance between them is short. Additionally, since the charging availability of a given EVCS influences the demand for near-distance and far-distance EVCS, it is necessary to analyze the data correlation between each set of EVCS nodes.

[0053] Furthermore, driving convenience for EV owners must be considered, as weather conditions can hinder their driving and the use of EVCS. External factors, such as holidays and weekends, can also influence travel purposes, distances, and EVCS demand. Therefore, to improve the accuracy of charging demand forecasting, the use of a single fixed adjacency matrix should be replaced with an approach that considers the mutual influence and mapping relationships between nodes and external factors when accounting for spatial dependencies.

[0054] Accordingly, in one embodiment of the present invention, in order to comprehensively consider factors affecting EVCS demand, a mutual adjacency matrix (MAM) can be constructed using static and dynamic attributes.

[0055] Static factors refer to static geographical information, such as the distance between nodes, and generally, strong correlations exist between nodes located close together. Dynamic factors can be used to represent characteristic correlations between nodes, EV user comfort, and external factors. As time-varying determinants of EVCS conditions, these can directly and indirectly influence EVCS demand.

[0056] According to one embodiment of the present invention, a MAM can be constructed using a distance matrix (static element) and three dynamic element matrices (feature correlation matrix, EV user comfort matrix, external factor matrix).

[0057] Below, we will explain the MAM configuration in more detail.

[0058] (1) Distance matrix based on static elements

[0059] The static attribute referred to as the distance between EVCS nodes can directly represent the correlation of EV charging demand, and a shorter distance indicates a stronger correlation. Location information for each node, such as longitude and latitude, can be obtained through GPS data. EVCS location information P=(lon l , lat l Based on ), the distance between EVCS nodes can be calculated using the Haversine formula as shown in Equation 1.

[0060]

[0061]

[0062] Here, R = 6371 km represents the radius of the Earth, and d is the distance between the x-th EVCS node and the y-th EVCS node, which can be calculated as d = θ x R. Based on Equations 1 and 2, the straight-line distance D between the two EVCS nodes. stl can be calculated. Once the distance between each EVCS node is calculated in this way, the distance matrix M can be used. DM It can be expressed as in mathematical formula 3.

[0063]

[0064] (2) Construction of dynamic feature correlation matrix, charging convenience matrix, and charging external factor matrix based on dynamic factors

[0065] Since there is a high or low correlation between charging demands of various EVCSs, it is important to effectively capture the characteristics of charging demand correlations between EVCS nodes. In one embodiment of the present invention, the charging demand correlation coefficient between various EVCS nodes can be calculated using the Pearson correlation coefficient method. This can be explained as follows.

[0066]

[0067] Here, x l wa y l represents the l-th feature of the two nodes, and σ x and σ y represents the standard deviation of the selected features. Dynamic feature correlation matrix M based on Pearson correlation coefficients FC It can be obtained as in mathematical formula 5.

[0068]

[0069] Furthermore, weather conditions can influence the charging demand of EV users. As time-varying determinants that determine driving conditions—such as road conditions and driver visibility—weather conditions can affect not only the use of EVs and EVCS but also actual travel intentions. For example, EVCS charging demand may decrease during inclement weather compared to clear weather.

[0070] Therefore, in one embodiment of the present invention, w UC (i,j) is defined in the range [0,1], and the better weather conditions in the current environment are w UC As the value of (i,j) approaches 1, conversely, the value may decrease under bad weather conditions. In particular, w UC(i,j) is used as a categorical vector distinguishing between clear, cloudy, heavily overcast, and light rain, as well as the presence of heavy rain, fog, thunder, and cyclone, and can represent EV user comfort values ​​under different weather conditions. Accordingly, the user charging convenience matrix M UC It can be obtained as shown in mathematical formula 6.

[0071]

[0072] In addition, external factors such as public holidays and days of the week (weekdays vs. weekends) can also affect EVCS charging demand. For example, EVCS demand on New Year's Day or weekends will be significantly higher than on weekdays due to differences in user purposes, such as commuting versus travel. EF The categorical category denoted by captures an external attribute corresponding to whether a specific day is a holiday or a weekend, which can be expressed as in Equation 7. This component factor matrix M EF It can be obtained as in mathematical formula 8.

[0073]

[0074]

[0075] Here, W D , W FC , W UC W EF represents the learnable weights of the matrix, and || and * represent connections (combinations) and convolutions, respectively. Also, W M wa b m is W M Represents the convolution output weights and bias terms.

[0076] In step 115, the computing device (100) applies the dynamic graph and the time series data to a relational graph convolution-based first deep learning model to generate spatial feature information.

[0077] The structure of a first deep learning model (mRGCN) according to one embodiment of the present invention is illustrated in FIG. 4.

[0078] Referring to FIG. 4, the first deep learning model can be configured to better capture local small changes in a matrix by combining graph convolution operations and residual elements. To this end, the first deep learning model can derive residual elements using a dynamic graph and the time series data.

[0079] The first deep learning model may include two graph convolution network blocks as shown in FIG. 4. In the graph convolution network blocks, graph data is processed in the spectral domain.

[0080] Therefore, the mutually adjacent matrix (W M Based on ), the normalized Laplacian matrix of the graph can be calculated as in Equation 10.

[0081]

[0082] Here, X l and X l+1 represents the EVCS eigenvalues ​​before and after graph convolution extraction, and And, And, D, I, and f represent the order, identity, and activation function, respectively.

[0083] Because the availability of EVCS is limited in smart energy communities, spatial dependencies of EVCS demand exist across a wide range. For example, the availability conditions of one EVCS can affect not only the EVCS status in surrounding areas but also the charging demand conditions of distant EVCSs. This indicates that there are two types of spatial dependencies in the spatial attributes of EVCS demand: local and global. To achieve better EVCS demand forecasting results, both local and global spatial dependencies of charging demand must be considered. As the layers of graph convolution become deeper, the ability to extract local dependencies decreases, which may weaken the ability to capture global spatial dependencies.

[0084] Accordingly, in one embodiment of the present invention, as shown in FIG. 4, a first deep learning model (mRGCN) having two layers of graph convolution network blocks can be configured.

[0085] In this way, by stacking two graph convolution network blocks, the first graph convolution network block can capture local spatial dependencies (local spatial feature values) based on dynamic graphs and time-series data (mutual adjacency matrices).

[0086] In addition, the second graph convolution network block can capture global spatial dependency (global spatial feature value) by learning similar EVCS demand patterns at near and far distances based on the results of the first graph convolution network block.

[0087] In this way, by stacking two layer graph convolution network blocks to form a first deep learning model, spatial dependencies (feature values) can be progressively extracted from a lower level to an upper level.

[0088] An activation block can be placed after the first and second graph convolution network blocks, respectively. It is a well-known fact that, despite the use of regularization and other techniques, introducing numerous activation functions into deeper neural networks as the network depth increases can degrade performance and negatively impact the model's training performance.

[0089] Accordingly, in one embodiment of the present invention, residual elements are applied to balance linear and non-linear transformations, and the prediction accuracy of the first deep learning model can be made not to include the deep structure of the network.

[0090] As illustrated in FIG. 4, the first deep learning model (mRGCN) can be composed of a combination of two graph convolution network blocks and a ReLU activation layer. Additionally, the first deep learning model (mRGCN) can generate final spatial feature information by combining residual elements with the outputs of the two graph convolution network blocks. This can be expressed mathematically as Equation 11.

[0091]

[0092] Here, R l+1 represents the final output of mRGN.

[0093] In step 120, the computing device (100) applies dynamic graphs and time series data to a second deep learning model to generate temporal feature information.

[0094] EVCS charging demand is influenced by charging conditions of the previous time along with multilayer spatial dependency. In one embodiment of the present invention, residual factors are applied to a first deep learning model (mRGCN), but only spatial relationships are processed and the temporal dependency of the EVCS charging demand sequence is not considered.

[0095] Accordingly, in one embodiment of the present invention, a complex time pattern considering the time dependency of EVCS charging-related data can be learned using a second deep learning model.

[0096] From a temporal perspective, changes in EVCS charging demand have several characteristics, such as periodicity, temporal regularity, and nonlinearity.

[0097] LSTM models outperform RNNs in handling sequence-based tasks involving long-term dependencies. They can capture long-term time dependencies of a dataset by storing long-term information from the collected dataset. For this reason, LSTMs are frequently adopted to handle time dependencies. Since EVCS charging demand has distinct time-series characteristics, LSTMs can be utilized to obtain important features of time-series variation from the acquired EVCS dataset.

[0098] Figure 5(a) shows the LSTM model structure. The LSTM model is an input gate i t , output gate o t , Oblivion Gate f t Includes, i t determines the information to preserve, and o t determines the result to be generated, and f t It can control the parts to be discarded. In FIG. 5(a), the LSTM model has a three-layer structure similar to the RNN model, but the hidden layer of the LSTM model has more units to manage information transfer. The mathematical formulas for calculating three gates and memory cells in each memory unit are given by Equations 12 to 16.

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] Here, W and b represent the weight matrix and bias vector for the gate, respectively, and σ represents the activation function expressed as a sigmoid function. Also, represents element-wise multiplication, and Tanh represents the hyperbolic tangent function. Based on the three gate functions of Equations 12 to 14 and the cell output state (Equation 15), the hidden layer output of Equation 16 can be derived.

[0105] Since LSTM models utilize only forward dependencies, they inevitably filter out important information due to long-term gated memory chains. Therefore, this can be resolved by using a Bi-LSTM model to combine the forward LSTM model and the backward LSTM model (see Figure 5(b)). The Bi-LSTM model improves the ability to explore long-term contextual dependencies in sequence prediction tasks and can generate more accurate sequence prediction results. EVCS charging demand exhibits strong periodicity and regularity; by utilizing the backward time dependency of the Bi-LSTM, periodic patterns of charging demand can be derived, and comprehensive predictions can be performed.

[0106] The Bi-LSTM model has two parallel LSTM layers, and the output can be represented as Equations 17 and 18.

[0107]

[0108]

[0109]

[0110] Here, LSTM FW and LSTM BW represents the forward LSTM and the back LSTM, respectively, and and is the time series input feature x learned through the Bi-LSTM model t It represents the hidden state of. Also, and and represent the weights of the forward and backward LSTMs, respectively, and the bias vector of the Bi-LSTM output layer is b y It is denoted as, and the final output of the Bi-LSTM model is b t It can be obtained as.

[0111] While LSTM and Bi-LSTM can achieve maximum predictability by learning temporal patterns through forward and backward propagation, EVCS charging demand depends on various dynamic external factors. Therefore, additional multilayer feedforward neural network prediction models can be used in conjunction with Bi-LSTM models to account for external factors such as time, weekday, past EVCS demand conditions, and average charging rates.

[0112] Referring to FIG. 6, a second deep learning model according to one embodiment of the present invention may include a bidirectional LSTM model and a multilayer feedforward neural network prediction model.

[0113] A bidirectional LSTM model can model the dependency on short-term charging demand (short-term charging demand pattern) using EVCS charging-related data (dynamic graph), and a multilayer feedforward neural network prediction model based on day-type trend feature information can model the demand pattern for time series data.

[0114] Time series samples X1={x t-n , 쪋, x n It is denoted as}, and X2 is denoted as X2={t, d, s, r}. Here, t, d, s, and r represent time, weekday, past EVCS demand status, and average charging rate, respectively. X1 and X2 represent the EVCS charging demand sequence history and the daily-type trend feature history, respectively.

[0115] The Bi-LSTM model uses X1 as input, and the output of the Bi-LSTM model can be combined with the output of an FNN model using X2 as input.

[0116] The output of the Bi-LSTM model is connected to the FNN model, and the FNN layer and output layer are linked, allowing for the output of each multi-stage advance charging demand forecast.

[0117] To summarize, by combining Bi-LSTM and FNN models, overall temporal dependency can be modeled through the Bi-LSTM model, and weekday type trend features can be processed through a multilayer FNN model.

[0118] Unlike conventional methods that consider only time-related EVCS demand characteristics through a circular network, according to one embodiment of the present invention, prediction accuracy can be improved by combining heterogeneous characteristics.

[0119] Final temporal feature output c t To obtain, the outputs of the Bi-LSTM model and the FNN layer can be connected, and then a dropout-applied FNN layer can be placed.

[0120]

[0121] Here, is the final output of the second deep learning model (c t Represents the weights, biases, and output of a multilayer feedforward neural network prediction model to generate ), where || represents the connection operator.

[0122] A charging demand prediction model according to one embodiment of the present invention is illustrated in FIG. 7. As illustrated in FIG. 7, the charging demand prediction model can perform target time step prediction by combining a first deep learning model, which is a relational graph convolution-based model, and a second deep learning model, which is a bidirectional LSTM-based model, in parallel.

[0123] To this end, as shown in Fig. 7, EVCS-related data consists of two forms: dynamic graphs and time series data, as described above.

[0124] By applying these dynamic graphs and time series data to the first deep learning model and the second deep learning model, respectively, spatial dependency (spatial feature information) and temporal dependency (temporal feature information) can be learned, respectively.

[0125] In step 125, the computing device (100) combines spatial feature information and temporal feature information, which are the outputs of the first deep learning model and the second deep learning model, to output hourly EVCS charging demand forecast values.

[0126] The outputs of the first and second deep learning models are R k and C m In this case, the final output of the charging demand forecasting model can be expressed as Equation 21.

[0127]

[0128] Here, k and m represent the number of hidden units in the final layers of the first deep learning model and the second deep learning model, respectively, and W st wa b st represents the learnable weights and bias parameters used to obtain the final prediction result (Y).

[0129] A method for operating an EVCS based on an EVCS charging demand forecasting model according to one embodiment of the present invention will be briefly described with reference to FIG. 8. For the convenience of understanding and explanation, the explanation will be based on CAISO.

[0130] In a liberalized retail electricity market, electricity users can manage electricity costs through various purchasing options. EVCS owners can participate in the power market as retailers and generate revenue by purchasing electricity and selling it to EV users. The primary objective of the EVCS optimal bidding process is to maximize the profits of EVCS owners to integrate EVs into the power system and increase the adoption of these vehicles.

[0131] In a Day-Ahead Market (DAM) according to one embodiment of the present invention, an observation station can maximize the profit of EVCS by submitting a 24-hour bid for the next day based on the hourly forecast value of aggregated EVCS charging demand obtained through an EVCS charging demand forecasting model.

[0132] When entering the actual operating day, the station must submit additional bids in the Real-Time Market (RTM) due to the underestimation of actual charging demand with EVCS, which incurs penalty costs for not fulfilling the DAM bid.

[0133] When actual charging power deviates from EVCS demand forecasts, this situation occurs, and a feedback loop may be formed between EV users and the energy market. For example, EV users tend to charge when real-time prices are low on actual charging days. Subsequently, based on actual charging demand, the Observatory submits additional increase real-time bids to the RTM, and real-time settlements may be announced every 15 minutes to meet charging demand when real-time prices are low.

[0134] In the CAISO energy market, DAM bids are binding, and market customers must consume their DAM bids the following day. Observatories can adjust DAM bids submitted to the RTM, but can only submit additional bids. If customers do not comply with the DAM bid results, they will be charged for costs and penalty fees.

[0135] In this market context, the observatory must first forecast the day-ahead (DA) power demand and submit a DAM power bid accordingly. Subsequently, at the RTM, the observatory adjusts its purchased power to ensure the charging demand of EVCS power and follows the DA bid. Therefore, the accuracy of EVCS charging demand forecasting plays a crucial role in the observatory's economic interests in the energy market.

[0136] An observatory may be an intermediary that predicts the charging demand of all EVCSs using a charging demand forecasting model according to one embodiment of the present invention and participates in the electric market on their behalf. It is assumed that such an observatory maximizes the profit of each EVCS by selling power to EVs and minimizes charging costs in both DAM and RTM.

[0137] The objective function for an observatory to participate in DAM can be expressed as maximizing profit as shown in Equation 22.

[0138]

[0139] Here, and represents the power sold to EV users, the DAM market bid, the power sold to EVCS at time h, and the DAM price at time h, respectively. Equation 22 aims to maximize the total observatory profit, including the revenue obtained from selling power to EV users and the cost of DAM power for the energy bid submitted to meet the EVCS charging demand.

[0140] In addition, in the second part of the optimization problem, the performance of the station on the bidding day can be evaluated using Equation 23. Although the station implements the proposed spatiotemporal prediction model in the Day Before (DA) stage, in reality, there is always a discrepancy between the predicted charge value and the actual charge value. Therefore, the station must adjust the power bidding to comply with the awarded DAM bid while meeting EVCS charge demand. Equation 23 aims to adjust the overall performance of the station in the power market and the role that EVCS charge demand forecasting plays.

[0141]

[0142]

[0143]

[0144] Here, h0 and ㅿH h represents the initial time of rolling horizon optimization and the length of the period, respectively. represent the additional power purchased in RTM, the power amount for which DAM bids were not fulfilled, and the power amount purchased in RTM, respectively. Price class and represent the power purchase price and the penalty price for failing to fulfill the DAM bid, respectively. Equation 23 includes the cost of purchasing additional power demand at RTM and the cost of failing to comply with the DAM bid, Equation 24 reflects the characteristics of the DAM bid, and Equation 25 represents the power quantity P Inc τ+ㅿH and P Pen τ+ㅿH It guarantees.

[0145]

[0146] FIG. 9 is a block diagram schematically illustrating the internal configuration of a computing device according to one embodiment of the present invention.

[0147] Referring to FIG. 9, a computing device (100) according to one embodiment of the present invention comprises a preprocessing unit (910), a learning unit (920), a prediction model unit (930), a memory (940), and a processor (950).

[0148] The preprocessing unit (910) is a means for dividing EVCS (Electric vehicle charging station) related data into dynamic graphs and time series data.

[0149] Since this is the same as described above in Figure 1, redundant explanations will be omitted.

[0150] The learning unit (920) is a means for training an EVCS charging demand prediction model using a learning data set (dynamic graphs, time series data, and correct answer data).

[0151] The prediction model unit (930) is a means for generating time-spatial feature information by applying dynamic graphs and time series data to an EVCS charging demand prediction model and using this to generate EVCS hourly charging demand prediction results.

[0152] More specifically, the prediction model unit (930) can generate spatial feature information by applying dynamic graphs and time series data to a relationship graph convolution-based first deep learning model of the EVCS charging demand prediction model. Next, the prediction model unit (930) can generate temporal feature information by applying dynamic graphs and time series data to a bidirectional LSTM-based second deep learning model. Next, the prediction model unit (930) can output an hourly charging demand prediction result of the EVCS by combining the spatial feature information and the temporal feature information.

[0153] A detailed explanation regarding this is the same as that described with reference to FIGS. 1 to 7, so a redundant explanation will be omitted.

[0154] The memory (940) stores various instructions (program codes) for performing a time-space graph convolution-based electric vehicle charging demand forecasting method according to one embodiment of the present invention.

[0155] The processor (950) is a means for controlling internal components (e.g., a preprocessing unit (910), a learning unit (920), a prediction model unit (930), a memory (940), etc.) of a computing device (100) according to one embodiment of the present invention.

[0156] An apparatus and method according to an embodiment of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the computer-readable medium may be those specifically designed and configured for the present invention, or they may be those known and available to a person skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0157] The hardware device described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

[0158] The present invention has been described above with reference to its embodiments. Those skilled in the art will understand that the present invention may be implemented in modified forms without departing from the essential characteristics of the invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the invention is defined by the claims, not by the foregoing description, and all variations within the scope of the claims should be interpreted as being included in the invention.

Claims

1. (a) A step of dividing EVCS (Electric vehicle charging station) related data into dynamic graphs and time series data; (b) a step of generating spatial feature information by applying the dynamic graph and the time series data to a relational graph convolution-based first deep learning model; (c) a step of generating temporal feature information by applying the dynamic graph and the time series data to a second deep learning model based on a bidirectional LSTM; and (d) A time-space graph convolution-based electric vehicle charging demand forecasting method comprising the step of combining the spatial feature information and the temporal feature information to output the hourly charging demand forecasting result of the EVCS.

2. In Paragraph 1, A time-space graph convolution-based electric vehicle charging demand forecasting method, characterized in that the dynamic graph above is a time series graph containing spatiotemporal correlations for each EVCS (Electric vehicle charging station) location and charging demand data for a continuous time series.

3. In Paragraph 2, The above time series data is, A time-space graph convolution-based electric vehicle charging demand forecasting method characterized by being a mutual adjacency matrix generated using static attributes based on the EVCS location and dynamic attributes affecting the charging demand of the EVCS.

4. In Paragraph 3, The above mutual adjacency matrix is, A time-space graph convolution-based electric vehicle charging demand forecasting method characterized by calculating the distance between EVCSs using the above-mentioned EVCS locations, constructing a distance matrix using the distance between EVCSs, constructing a dynamic feature correlation matrix based on the charging demand correlation with adjacent EVCSs based on Pearson correlation coefficients, constructing a charging convenience matrix based on weather information, constructing a charging external factor matrix based on weekday information regarding holidays and commuting, and then combining the distance matrix, the dynamic feature correlation matrix, the charging convenience matrix, and the charging external factor matrix.

5. In Paragraph 3, The above-mentioned first deep learning model is, A time-space graph convolution-based electric vehicle charging demand forecasting method characterized by deriving residual elements using the time series graph and the mutual adjacency matrix, and generating spatial feature information by adding the residual elements to the result of applying the time series graph and the mutual adjacency matrix to a relational graph convolution-based model.

6. In Paragraph 5, The above-mentioned first deep learning model is, A first graph convolution network block that receives the time series graph and the mutual adjacency matrix as input and extracts local spatial feature values ​​for local spatial dependencies reflecting the charging demand patterns of EVCSs located at adjacent distances; and It includes a second graph convolution network block that extracts global spatial feature values ​​for global spatial dependencies reflecting the status and charging demand patterns of EVCSs located at a distance using the local spatial feature values ​​which are the outputs of the first graph convolution network block, wherein A time-space graph convolution-based electric vehicle charging demand forecasting method characterized by generating spatial feature information by adding the residual element to the global spatial feature value.

7. In Paragraph 1, A time-space graph convolution-based electric vehicle charging demand forecasting method characterized in that the above time series data is a sequence of day-type trend features including time of day, day of week, past charging demand status and average charging rate.

8. In Paragraph 2, The above second deep learning model is, A multilayer feedforward neural network model that receives the above time series data as input and generates an output vector reflecting the EVCS demand pattern; and It includes a bidirectional LSTM model that receives the above time series graph as input and outputs a temporal feature vector, A time-space graph convolution-based electric vehicle charging demand forecasting method characterized by generating temporal feature information by combining an output vector, which is the output of the above-described multilayer feedforward neural network model, and a temporal feature vector, which is the output of the above-described bidirectional LSTM model.

9. A computer-readable recording medium having program code for performing the method according to claim 1.

10. A preprocessing unit that divides EVCS (Electric vehicle charging station) related data into dynamic graphs and time series data; and A computing device comprising a prediction model unit that generates spatial feature information by applying the dynamic graph and the time series data to a relationship graph convolution-based first deep learning model, generates temporal feature information by applying the dynamic graph and the time series data to a bidirectional LSTM-based second deep learning model, and then outputs an hourly charging demand prediction result for an EVCS by combining the spatial feature information and the temporal feature information.

11. In Paragraph 10, A computing device characterized in that the above dynamic graph is a time series graph containing spatiotemporal correlations for each EVCS (Electric vehicle charging station) location and charging demand data for a continuous time series.

12. In Paragraph 11, The above time series data is, A computing device characterized by being a mutually adjacent matrix using static attributes based on the above EVCS location and dynamic attributes that affect the charging demand of the EVCS.

13. In Paragraph 12, The above preprocessing unit is, A computing device characterized by calculating the distance between EVCSs using the above-mentioned EVCS locations, constructing a distance matrix using the distance between the above-mentioned EVCSs, constructing a dynamic feature correlation matrix based on the charging demand correlation with adjacent EVCSs based on Pearson correlation coefficients, constructing a charging convenience matrix based on weather information, and constructing a charging external factor matrix based on day information regarding holidays and commuting, and then combining the distance matrix, the dynamic feature correlation matrix, the charging convenience matrix, and the charging external factor matrix to generate the above-mentioned mutual adjacency matrix.

14. In Paragraph 12, The above-mentioned first deep learning model is, A computing device characterized by deriving residual elements using the time series graph and the mutual adjacency matrix, and generating spatial feature information by adding the residual elements to the result of applying the time series graph and the mutual adjacency matrix to a relational graph convolution-based model.

15. In Paragraph 14, The above-mentioned first deep learning model is, A first graph convolution network block that receives the time series graph and the mutual adjacency matrix as input and extracts local spatial feature values ​​for local spatial dependencies reflecting the charging demand patterns of EVCSs located at adjacent distances; and It includes a second graph convolution network block that extracts global spatial feature values ​​for global spatial dependencies reflecting the status and charging demand patterns of EVCSs located at a distance using the local spatial feature values ​​which are the outputs of the first graph convolution network block, wherein A computing device characterized by generating spatial feature information by adding the residual element to the global spatial feature value.

16. In Paragraph 11, The above second deep learning model is, A multilayer feedforward neural network model that receives the above time series data as input and generates an output vector reflecting the EVCS demand pattern; and It includes a bidirectional LSTM model that receives the above time series graph as input and outputs a temporal feature vector, A computing device characterized by generating temporal feature information by combining an output vector, which is the output of the above-described multilayer feedforward neural network model, and a temporal feature vector, which is the output of the above-described bidirectional LSTM model.

Images