Electric vehicle charging load prediction method based on multi-source spatio-temporal data and user behaviors

Abstract

The invention discloses an electric vehicle charging load prediction method based on multi-source spatio-temporal data and user behaviors, and the method comprises the steps: randomly generating an initial node of a user in a certain functional region based on a trip chain in a certain scene, determining a destination functional region according to an OD matrix in the scene, randomly generating a destination node in the destination functional region, and predicting the charging load of the user. Based on an improved Dijkstra path planning algorithm of Pareto optimization, performing path selection according to a user tag; at a destination node, randomly generating an end point node in an end point functional area, selecting a path according to a user label, and driving to the end point node; and in the process of going to the next node, if a charging demand exists, determining a nearby charging station which can be reached by the residual SOC, selecting the charging station with the lowest comprehensive cost, determining a time-economy double-constraint charging mode after the electric vehicle arrives at the charging station from the optimal path, and recording charging load data by the charging station for the subsequent constant volume site selection of the electric vehicle charging station.

Description

Technical Field

[0001] The present invention belongs to the field of electric vehicle charging load prediction, and in particular relates to an electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior. Background Art

[0002] Against the backdrop of growing global environmental awareness and the mainstreaming of sustainable energy development, the number of electric vehicles (EVs) continues to rise rapidly due to their clean and efficient nature. While the widespread adoption of EVs brings significant environmental benefits, it also poses numerous challenges to power system planning and operation. Due to the highly random nature of EV charging behavior and their uneven temporal and spatial distribution, the disorderly charging of large numbers of EVs can lead to voltage drops at local nodes in the distribution network and line overloads, severely impacting the safe and stable operation of the power system. Therefore, accurately predicting EV charging load is crucial for rational power system planning, optimized scheduling, and improved power supply reliability.

[0003] Previous methods for forecasting electric vehicle charging loads have relied on historical load data, employing models such as time series analysis and regression analysis. However, these methods often overlook the diverse nature of electric vehicle user behavior and the role of spatiotemporal factors, resulting in forecasting accuracy that fails to meet practical needs. In recent years, with the rapid development of information technology, the acquisition of multi-source spatiotemporal data has become increasingly accessible. For example, smart meters can provide real-time user electricity usage data, the Global Positioning System (GPS) and Geographic Information System (GIS) can be used to obtain information on electric vehicle locations and driving trajectories, and social media platforms can be used to mine behavioral data such as user travel habits and charging preferences. This multi-source spatiotemporal data has created new opportunities for in-depth research into electric vehicle charging behavior and improved charging load forecasting accuracy. Summary of the Invention

[0004] In view of this, the present invention discloses a method for predicting electric vehicle charging load based on multi-source spatiotemporal data and user behavior.

[0005] The present invention achieves the above technical objectives through the following technical means.

[0006] Electric vehicle charging load forecasting method based on multi-source spatiotemporal data and user behavior:

[0007] Import a map of a certain area into QGIS software, select the prediction area on the map, determine the road network topology within the prediction area, and divide the selected prediction area into functional areas according to POI data;

[0008] Obtain user feature data, perform adaptive clustering algorithm analysis on user features based on driving style, time preference, and energy efficiency preference, and label the user;

[0009] Solve the OD matrix under different scenarios and set up travel chains;

[0010] In a certain scenario, based on the travel chain, the user's initial node is randomly generated in a certain functional area, the destination functional area is determined according to the OD matrix in the scenario, the destination node is randomly generated in the destination functional area, and the improved Dijkstra path planning algorithm based on Pareto optimization is used to select a path based on the user's label; when arriving at the destination node, the terminal node is randomly generated in the terminal functional area, and the improved Dijkstra path planning algorithm is used to select a path based on the user's label, and the user travels to the terminal node;

[0011] In the process of heading to the next node, if there is a demand for charging, the remaining SOC of the electric vehicle determines the nearby charging stations that can be reached. A charging station selection model that considers the comprehensive cost is designed. The user selects the charging station with the lowest comprehensive cost among the nearby charging stations that can be reached. The improved Dijkstra path planning algorithm determines the optimal path to the charging station with the lowest comprehensive cost. After arriving at the charging station, the charging method is determined through a time-economy dual-constraint strategy. The charging station records the charging load data for subsequent fixed-capacity site selection of electric vehicle charging stations.

[0012] Furthermore, the improved Dijkstra path planning algorithm based on Pareto optimization specifically includes:

[0013] The topological information of the road network is obtained from node data, road grade, road length, and adjacency matrix, and the path weight is defined in multiple dimensions. Each edge corresponds to a five-tuple (t uv ,e uv ,c uv ,r uv ,s uv );

[0014] Establish a priority queue. The element format of the priority queue is: (node ​​number, comprehensive score F(v), predecessor node). The predecessor node is used to record the path source to facilitate backtracking the complete path. Before path planning begins, the initial node is added to the priority queue. The formula for the comprehensive score is:

[0015] F(v)=ω t ·t uv +ω e ·e uv +ω c c uv +ω r ·(1-r uv )+ω s ·s uv

[0016] Among them, (ωt ,ω e ,ω c ,ω r ,ω s ) is the weight after dynamic adjustment of reinforcement learning, t uv is the travel time, e uv is energy consumption, c uv is carbon emissions, r uv For road comfort, s uv is the traffic congestion index;

[0017] Select the node u with the smallest F(v) from the priority queue for expansion, and remove u from the priority queue. Obtain all adjacent edges (u, v) of node u based on the adjacency matrix. For each edge (u, v), calculate the new comprehensive score F′(v) of node v:

[0018] F′(v)=F(u)+ω t ·t uv +ω e ·e uv +ω c c uv +ω r ·(1-r uv )+ω s ·s uv

[0019] If v is not in the priority queue, add v to the priority queue and record its comprehensive score F′(v); if v is in the priority queue and F′(v) < F(v), then update F(v) = F′(v) and adjust the priority queue order to ensure that the queue is always arranged in ascending order of comprehensive score; this process is repeated until the priority queue is empty;

[0020] During the priority queue expansion process, a set of candidate paths is generated. For example, two candidate paths have solution indices of (T2, E2, C2, R2, S2) and (T1, E1, C1, R1, S1). The indices of the two solutions are determined according to the following rules:

[0021] If T1≤T2, E1≤E2, C1≤C2, R1≥R2, S1≤S2, and at least one inequality strictly holds, then solution 1 Pareto dominates solution 2, solution 2 is discarded, and solution 1 is directly output as the result of the improved Dijkstra path planning algorithm;

[0022] If T2≤T1, E2≤E1, C2≤C1, R2≥R1, S2≤S1, and at least one inequality strictly holds, then solution 2 Pareto dominates solution 1, solution 1 is discarded, and solution 2 is directly output as the result of the improved Dijkstra path planning algorithm;

[0023] If the above two conditions are not met, the two solutions do not dominate each other, and the Q-learning reinforcement learning algorithm is introduced to achieve the target weight (ω t ,ω e ,ω c ,ω r ,ω s ) dynamic optimization. Through continuous iterative learning, the algorithm automatically adapts to user preferences and scenario changes, updates the weights in the comprehensive score formula, and then affects the node expansion priority of the Dijkstra algorithm, starting again from the establishment of the priority queue;

[0024] Among them, T1, E1, C1, R1, and S1 are the driving time, energy consumption, carbon emissions, road comfort, and traffic congestion index of the first candidate path, respectively; T2, E2, C2, R2, and S2 are the driving time, energy consumption, carbon emissions, road comfort, and traffic congestion index of the second candidate path, respectively.

[0025] Furthermore, the travel time t uv Obtained by the improved road resistance function, the improved road resistance function is:

[0026] T=T L +T J

[0027]

[0028] f(T temp )=1+0.005·|T temp -20|T temp

[0029] Among them, T is the improved road resistance function, T L is the section impedance, T J is the node impedance, t0 is the free travel time of the road section, y′ is the corrected saturation, Q is the road flow, C is the road capacity, λ level Road grade adjustment factor, f(T temp ) is the temperature correction function, λ habit is the driving habit coefficient, k is the signal period, τ is the green light signal ratio, y is the node saturation, λ turn is the steering influencing factor, α is the coefficient of the corrected saturation linear term, β is the coefficient of the corrected saturation quadratic term, q is the node impedance calculation parameter, and γ is the corrected saturation adjustment factor.

[0030] Furthermore, the energy consumption e uv Obtained from the improved energy consumption model, the improved energy consumption model is:

[0031]

[0032]

[0033] Among them, E base is the basic energy consumption per unit mileage of the vehicle under ideal conditions, E speed is the energy consumption term affected by speed, E ac is the additional energy consumption for the interior temperature, C B (T) is the battery capacity under the influence of temperature, C t is the actual capacity of the vehicle, δ battery is the battery temperature sensitivity coefficient, k habit is the static coefficient of the original driving habit, f aggressive is the frequency of sudden acceleration and braking obtained by the vehicle sensor, Δk is the fluctuation coefficient, P c is the cooling power of the air conditioner, P h is the air conditioning heating power, T Lab is the road impedance on road section ab, i.e., the travel time, l ab is the length of section ab, v ab is the actual vehicle speed on road section ab.

[0034] Furthermore, the charging station selection model considering comprehensive costs is specifically as follows:

[0035] C t' =ζ·C o +(1-ζ)·C p

[0036] C o =C time +C cost +C queue +C loss

[0037] C time =ω1·t

[0038] C cost =p·E charge

[0039] C queue =ω2·t queue

[0040] C loss =ω3·ΔE loss

[0041] C p =C f +C s +C q

[0042]

[0043] Among them, C t' is the total decision cost, C o is the objective cost, C p is the psychological cost, ζ is the weight, C time is the time cost, C cost is the charging cost, C queue is the queuing cost, C loss is the energy cost, C f is the familiarity effect, C s For the quick gratification effect, C q is the loss aversion effect, ω1 is the user's time cost coefficient, p is the electricity price of the charging station, E charge is the estimated charge capacity, t queue is the queuing time at the charging station, ω2 is the queuing time cost coefficient, ΔE loss To calculate the energy consumption of driving to the charging station, ω3 is the energy cost coefficient, λ f is the basic cost of familiarity, I info is the information completeness index, t charge is the actual charging time, t ref is the reference charging time, λ ins is the psychological cost coefficient of instant satisfaction, n q is the real-time queue number, λ q is the psychological cost coefficient of queuing.

[0044] Furthermore, the remaining SOC of the electric vehicle is used to determine the nearby charging stations that the electric vehicle can reach, including:

[0045]

[0046] Among them, S OC is the current power ratio, S dest S is the proportion of power consumption when traveling to the next destination. safe is the safe power threshold, P cha is the charging probability; and E a'b' is the energy consumption of traveling from the current location a' to the next destination b', C t is the actual capacity of the vehicle;

[0047] When the current power ratio is less than the safe power threshold plus the power ratio corresponding to the expected power consumption to the next destination, that is, S OC <S safe +S dest , at this time the charging probability P cha If it is judged as 1, it is considered that there is a need for charging; when the current power ratio is less than 20%, it is judged that there is a need for charging. cha Also 1.

[0048] Furthermore, the charging method is determined by the strategy of dual constraints of time and economy, specifically:

[0049] Considering the rated power of the charging equipment and the actual time available for charging, the theoretical charging power P is constructed. EV =P km ×0.9×t available , where P km is the charging power, t available is the rechargeable time, and t available =t stay -t queue , t stay is the terminal residence time, t queue queue times for charging stations;

[0050] Limited by the upper limit of battery capacity, the target power ratio Among them, SOC now is the current battery SOC, C t is the actual capacity of the vehicle;

[0051] Calculate the actual charge capacity ΔE actual :ΔE actual =C t ×(SOC target -SOC now ), calculate the charging cost C charge :C charge =ΔE actual ×p, where p is the electricity price of the charging station;

[0052] Determine the target power ratio SOC of fast charging and slow charging target Is it greater than SOC? min If it is greater than, then choose the charging cost C of fast charging and slow charging. charge If the lower option is less than , then select the target power ratio SOC of fast charging and slow charging. target The higher option.

[0053] Furthermore, the user characteristics are analyzed using an adaptive clustering algorithm according to driving style, time preference, and energy efficiency preference, and the user is labeled, specifically:

[0054] The original data set is obtained from traffic survey data, and relevant data of multiple brands and models commonly seen in the market are imported. Then, the data is filtered and cleaned to determine the characteristic data to distinguish driving style, time preference, and energy efficiency preference. The user's driving style is distinguished as aggressive or gentle according to mileage, acceleration from 0 to 100 km / h, and dwell time at the destination. The user's preference is distinguished as high time priority or low time priority according to the start time of the trip. The user's preference is distinguished as high energy efficiency priority or low energy efficiency priority according to the age of the vehicle, energy consumption per 100 kilometers, and fast charging power.

[0055] The density peak clustering algorithm is used to automatically identify the cluster center by calculating the local density and relative distance of user feature data points, thereby achieving effective clustering of user features.

[0056] Furthermore, the local density ρ of all user feature data points is calculated by the “density-distance” decision graph. i' and the relative distance δ i’ Sort them from large to small and select ρ i' and δ i' The two points that are both in the top 5% are used as cluster centers, and each user feature data point is assigned to the cluster center closest to it to obtain the cluster label.

[0057] Furthermore, the functional areas are divided into residential areas, social and entertainment areas, work areas, commercial areas, restaurants, hospitals, schools, and others.

[0058] The beneficial effects of the present invention are:

[0059] (1) This invention discloses a method for predicting electric vehicle charging load based on multi-source spatiotemporal data and user behavior. By using POI data to divide the prediction area, obtain road network data, and perform adaptive clustering algorithm analysis on user characteristics, it can effectively extract user travel data characteristics. Compared with traditional prediction methods based on a single data source, it can more comprehensively express the impact of various variables on load. Clustering can obtain the distribution characteristics of vehicles and charging when different groups of people drive electric vehicles, making the prediction results more consistent with real-world scenarios.

[0060] (2) The present invention considers various factors such as road grade, air conditioning use, and driving habits to construct an improved road resistance function model and an improved energy consumption model, while solving the OD matrix under different scenarios and setting up travel chains. An improved Dijkstra path planning algorithm based on multi-source dynamic information fusion and reinforcement learning is proposed, which automatically selects paths based on the user's preferences for time and energy consumption. The above process comprehensively reflects the user's differentiated travel preferences and the impact of actual road conditions, making the prediction process more suitable for complex travel scenarios and able to more accurately predict load conditions, thereby effectively reducing the operating costs of the power grid and alleviating the load pressure on the power grid.

[0061] (3) This invention improves user demand judgment, ensuring that the remaining SOC can reach nearby charging stations. It also designs a charging station selection model based on dynamic fuzzy reasoning and user psychology, comprehensively considering traditional objective costs and psychological costs. In addition, it introduces a charging method decision system based on deep reinforcement learning, which infers and judges user charging methods based on multiple influencing factors. Combined with Monte Carlo simulation, it can achieve effective prediction of regional electric vehicle charging load. These measures can more effectively predict the actual demand for regional charging stations and provide strong support for charging facility planning and operation management. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for predicting electric vehicle charging load based on multi-source spatiotemporal data and user behavior according to the present invention;

[0063] Figure 2 To predict the regional road network map;

[0064] Figure 3(a) shows the driving style clustering decision diagram;

[0065] Figure 3(b) shows the 3D visualization of the driving style clustering results;

[0066] Figure 4(a) is the time preference clustering decision diagram;

[0067] Figure 4(b) is the time preference distribution histogram;

[0068] Figure 5(a) is the energy efficiency preference clustering decision diagram;

[0069] Figure 5(b) shows the energy efficiency preference clustering result;

[0070] Figure 6(a) is a Gaussian fitting graph of the travel time of users with aggressive driving style;

[0071] Figure 6(b) shows the Gaussian fitting graph of the travel time of users with a mild driving style;

[0072] Figure 6(c) shows the Gaussian fitting graph of the terminal dwell time of the aggressive driving style;

[0073] Figure 6(d) shows the Gaussian fitting graph of the terminal dwell time for the mild driving style;

[0074] Figure 7 This is a flowchart of the improved Dijkstra path planning algorithm based on multi-source dynamic information fusion and reinforcement learning;

[0075] Figure 8 Flowchart for selecting charging stations considering comprehensive costs. DETAILED DESCRIPTION

[0076] The following further illustrates and explains an electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior in conjunction with the accompanying drawings and embodiments.

[0077] As attached Figure 1 As shown, a method for predicting electric vehicle charging load based on multi-source spatiotemporal data and user behavior in this embodiment includes the following steps:

[0078] Step A: Divide the prediction area into 8 areas using POI data, extract node data, road grade, road length, and adjacency matrix; it includes the following steps:

[0079] Step A1: Import the map of a certain area of ​​a prefecture-level city (in this example, the Jiangning District of Nanjing City) into QGIS software, select the prediction area on the map, and determine the road network topology within the prediction area, as shown in the attached figure. Figure 2 The selected prediction area is divided into functional areas according to POI data, specifically residential area (R), social and entertainment area (S), work area (W), commercial area (P), catering (E), hospital (H), school (C), and others (O).

[0080] Step A2: Extract node data, road grade, road length, and adjacency matrix from the road network.

[0081] Step B: Obtain user feature data, perform adaptive clustering algorithm analysis on user features based on driving style, time preference, and energy efficiency preference, and label the users; use a Gaussian mixture model to fit user-related travel features for aggressive and moderate driving, and perform accuracy analysis on the fitting results; this includes the following steps:

[0082] Step B1: Obtain a raw data set from the traffic survey data and import relevant data from multiple common brands and models in the market. This data is then filtered and cleaned to obtain the following feature data: vehicle age (years), trip start time (minutes), mileage (miles), final dwell time (minutes), 0-100km / h acceleration (seconds), energy consumption per 100km / h (kWh / 100 miles), fast-charging power (kW), and battery capacity (kWh). This feature data is used to distinguish between aggressive and gentle driving styles based on mileage, 0-100km / h acceleration, and final dwell time. The user preference for high or low time priority is distinguished based on trip start time. Furthermore, the user preference for high or low energy efficiency is distinguished based on vehicle age, energy consumption per 100km, and fast-charging power.

[0083] Normalize the feature data and use minimum-maximum normalization to eliminate the dimension effect. The formula is:

[0084]

[0085] Among them, x ij is the jth feature value of the i-th user, x′ ij is the normalized eigenvalue.

[0086] The density peak clustering algorithm is used to automatically identify the cluster center by calculating the local density and relative distance of user feature data points, thereby achieving effective clustering of user features.

[0087] First, cluster the driving styles: extract the three features of mileage, acceleration from zero to hundred kilometers, and dwell time at the destination. For n user feature data points, calculate the Euclidean distance d between any two users i' and j'. i'j' :

[0088]

[0089] Where x′ i1 , x′ i2 , x′ i3 , x′ j1 , x′ j2 , x′ j3 are the normalized values ​​of mileage, acceleration at zero point, and dwell time at the terminal of users i' and j', respectively.

[0090] Sort all Euclidean distances from small to large, and take the distance value corresponding to the 2% quantile (2% of the data is less than or equal to this value) as the cutoff distance d c .

[0091] The truncated kernel function is used to calculate the local density of users i' and j'. The calculation formula is as follows:

[0092]

[0093] Among them, ρ i' is the local density, χ(x) is the cutoff function, when x < 0, χ(x) takes 1, when x ≥ 0, χ(x) takes 0.

[0094] The relative distance between users i' and j' is calculated as follows:

[0095]

[0096] Relative distance δ i’ It represents the minimum distance from a certain user feature data point i' to a user feature data point with a higher local density. If a certain user feature data point is the feature data point with the highest local density, then δ i’ Defined as the maximum distance from the user feature data point to all other user feature data points.

[0097] Draw a “density-distance” decision diagram and calculate the local density ρ of all user feature data points i' and the relative distance δ i’ Sort them from large to small and select ρ i' and δ i' The two points that are both in the top 5% are used as cluster centers. Each user feature data point is assigned to the cluster center with the closest distance to obtain a cluster label. Based on the clustering results, driving styles can be divided into "aggressive driving" and "gentle driving."

[0098] Differentiating time preferences based on the start time of the trip and differentiating energy efficiency preferences based on vehicle age, energy consumption per 100 kilometers, and fast charging power are consistent with the above steps.

[0099] Through the above three clustering, each user is given three different dimensional labels, representing their driving style (aggressive driving or gentle driving), time preference (high time priority or low time priority) and energy efficiency preference (high energy efficiency priority or low energy efficiency priority). The clustering results are shown in the attached figure. Figure 3(a) 、 3(b) , 4(a), 4(b), 5(a), 5(b).

[0100] For each clustering result, the silhouette score is used to evaluate the clustering quality. The silhouette score ranges from -1 to 1, with values ​​closer to 1 indicating better clustering results. The calculation formula is as follows:

[0101]

[0102] Among them, a H is the average distance from the feature data point H to other feature data points in the same cluster, b H It is the average distance from the feature data point H to the nearest different cluster.

[0103] Step B2: Select the trip start time and the destination dwell time as travel characteristics. These characteristics can better reflect the user's travel habits and behavior patterns. After filtering and cleaning all users, the data set is divided into an aggressive driving user data set and a moderate driving user data set based on the clustering results. Use the Gaussian mixture model to fit the two travel characteristics in the two data sets to obtain the travel characteristic curve.

[0104] The probability density function of the Gaussian mixture model can be expressed as:

[0105]

[0106] Where K is the number of Gaussian components, π k is the weight of the kth Gaussian component, N(x|μ k ,∑ k) is the probability density function of the k-th Gaussian distribution, which is in the form of:

[0107]

[0108] Among them, μ k is the mean, Σ k is the covariance matrix.

[0109] Randomly initialize the weight π of each Gaussian component k , mean μ k and the covariance matrix Σ k ; Calculate the characteristic data point x corresponding to each travel feature H The posterior probability γ belonging to the kth Gaussian component Hk , the formula is:

[0110]

[0111] According to the posterior probability γ Hk Update weights, mean, and covariance matrix. When the update amount of the three parameters of weights, mean, and covariance matrix is ​​less than 1×10 -4 Or when the maximum number of iterations, 100, is reached, the iteration is stopped.

[0112] Taking the trip start time of the intense driving user data set as an example, determine its minimum value x in the feature data min and the maximum value x max , in [x min ,x max ]Generate a series of uniformly distributed discrete sampling points x1, x2, ..., x m , each sampling point x i Substitute the probability density function p(x) of the Gaussian mixture model and calculate its probability density value. i is the horizontal axis, and the corresponding probability density value p(x i ) is the vertical coordinate, and all points (x i ,p(x i )) are connected into a smooth curve, which is the fitting curve of the trip start time.

[0113] The Gaussian curve fitting steps for the dwell time at the end of the aggressive and moderate driving user datasets were consistent with the Gaussian curve fitting steps for the trip start time of the aggressive driving user dataset. The Gaussian curve fitting results are shown in Figures 6(a), (b), (c), and (d).

[0114] The mean square error (MSE) and the coefficient of determination (R 2 ) to evaluate the fitting results of relevant travel characteristic curves. The mean square error (MSE) and the coefficient of determination (R2 )The formula is as follows:

[0115]

[0116] Among them, y H is the true value, is the fitted value, n is the number of samples, is the mean of the true values, R 2 The value range is [0, 1].

[0117] Step C: Considering factors such as season, weather, and time of day, obtain the road traffic volume on the predicted area path; considering factors such as road grade, temperature, and driving habits, construct an improved road resistance function model; this includes the following steps:

[0118] Step C1: Divide a day into six periods, each lasting four hours. Simplify the weather classification into sunny and rainy days, and combine the four seasons to form 48 subdivided scenarios. Obtain road traffic on regional routes under various scenarios through Baidu Maps.

[0119] Step C2: Divide the road resistance function into a road resistance function and a node resistance function, and construct an improved road resistance function that takes into account factors such as road grade, temperature, and driving habits. The formula for the improved road resistance function is:

[0120] T=T L +T J

[0121]

[0122] f(T temp )=1+0.005·|T temp -20|T temp

[0123] Among them, T is the improved road resistance function, T L is the section impedance, T J is the node impedance, t0 is the free travel time of the road section, y′ is the corrected saturation, Q is the road flow, C is the road capacity, λ level Road grade adjustment factor, f(T temp ) is the temperature correction function, λ habit is the driving habit coefficient, k is the signal period, τ is the green light signal ratio, y is the node saturation, λ turn is the steering influencing factor, α is the coefficient of the corrected saturation linear term, β is the coefficient of the corrected saturation quadratic term, q is the node impedance calculation parameter, and γ is the corrected saturation adjustment factor.

[0124] Obtain the road section length L (unit: km) from step A2, design the zero flow velocity v (unit: km / h) based on the road grade, and calculate t0 using the following formula:

[0125]

[0126] Set the road grade adjustment factor, main road λ level =1.0, secondary road λ level =1.2; Get real-time temperature T temp ;λ habit According to the driving style clustering result of step B1, the aggressive driving habit =1.1, gentle driving λ habit = 0.9; γ is determined by fitting the degree of influence of the signal cycle on y′, and is set to 0.1-0.3; k is the time it takes for the signal light to complete one cycle, and is usually in the range of 60-240 seconds; α is determined by fitting traffic data, and is in the range of 0.1-0.5; β is determined by fitting traffic data, and is in the range of 0.05-0.2; the road flow Q is obtained from step C1, and the corrected saturation y′ is calculated in combination with the road capacity C (set according to the road grade); the above parameters are substituted into the formula to solve T L .

[0127] Set the steering influence factor λ turn , turn left to set λ turn =1.5, turn right to set λ turn =1.2, set λ for straight line turn =1.0; τ is set according to the traffic flow at the intersection. For busy intersections, the green light time accounts for a high proportion, so τ is set to 0.5. For small intersections, τ is set to 0.4. Combined with the traffic flow at the node, q is set to a range of 0.5-1.5. Add the traffic volume in all directions at the node to obtain the total node traffic volume Q node ; Determine the lane saturation flow rate based on the main road and secondary road. The saturation flow rate of each lane on the main road is S main Set to 1800 vehicles / hour, the saturation flow rate of each lane of the secondary road is S sub Set to 1600 vehicles / hour; Combine the signal cycle k and the green light signal ratio τ to obtain the effective green light time kτ; The node traffic capacity C node The formula is:

[0128]

[0129] Among them, n main is the total number of main road import lanes at the node, n sub is the total number of entrance lanes of the secondary road at the node.

[0130] According to the total node traffic volume Q node and node capacity C node , calculate y, the specific formula is as follows:

[0131]

[0132] Substitute the formula to calculate T J , the section impedance T L and node impedance T J Add them together to calculate the total resistance function T.

[0133] Step D: Develop an improved energy consumption model taking into account air conditioning usage, battery capacity, and driving habits. The specific steps are as follows:

[0134] An improved energy consumption model is established based on the traditional energy consumption model by considering the effects of air conditioning use, battery capacity, and driving habits. The formula of the improved energy consumption model is:

[0135]

[0136] Among them, E base is the basic energy consumption per unit mileage of the vehicle under ideal conditions, obtained from the energy consumption per 100 kilometers data in step B1; E speed is the energy consumption item affected by speed; E ac The additional energy consumption for the vehicle interior temperature is: when the temperature is above 30°C, the air conditioner is used for cooling, and when it is below 5°C, it is used for heating; ab is the length of section ab, obtained from step A2; v ab is the actual vehicle speed on road section ab; Pc is the air conditioning cooling power; P h is the air conditioning heating power; T Lab is the road impedance on road section ab, i.e., the travel time; battery is the battery temperature sensitivity coefficient. The further the temperature deviates from 25°C, the smaller the denominator value is and the greater the energy consumption E is. This is consistent with the physical characteristics of battery charging and discharging efficiency fluctuating with temperature. Combining the physical characteristics of the battery and common empirical values, δ battery Take 0.005 / ℃; k habit is the static coefficient of the original driving habit (0.9 for gentle driving and 1.1 for intense driving), reflecting the difference in basic energy consumption for different driving styles; aggressive is the frequency of sudden acceleration and braking obtained through vehicle sensors; Δk is the fluctuation coefficient, which dynamically amplifies the impact of intense driving behavior on energy consumption when driving is intense and sudden acceleration is frequent, making the model more suitable for actual driving scenarios. The value range is 0-0.5.

[0137] C B (T) is the battery capacity under the influence of temperature:

[0138]

[0139] Among them, C t is the actual capacity of the vehicle, obtained from the battery capacity data in step B1; K p is the temperature coefficient, which is 0.01 / ℃.

[0140] Step E: Solve the OD matrix under different scenarios and set up the travel chain; the specific steps are as follows:

[0141] Based on the urban functional zoning results obtained in step A1, the scope and distribution of different functional zones, such as residential, commercial, work, and entertainment areas, are clarified. Based on the season, weather, and time of day considered in step C1, 48 segmented scenarios are established. The travel data for each scenario is traversed. For each trip record, if the departure functional zone is numbered i' and the destination functional zone is numbered j', the value of the element in the i'th row and j'th column of the OD matrix is ​​incremented by 1.

[0142] For each scenario’s OD matrix, calculate the sum of the elements in each row, and the sum of the elements in the i’th row S i” represents the total number of trips from functional area i”, and the formula is:

[0143]

[0144] Among them, M i”j” is the value of the element in the i'th row and j'th column in the OD matrix.

[0145] The element P in the i'th row and j'th column of the OD matrix i”j” The calculation formula is:

[0146]

[0147] This value represents the probability of transitioning from functional area i' to functional area j'.

[0148] Through the above calculations, the OD matrix in different scenarios is obtained.

[0149] Set up travel chains and divide them into simple chains and complex chains. Simple chain setting: Identify all residential areas from the functional area data, and for each residential area, according to its corresponding OD matrix, calculate the probability P i”j” Based on this, a non-residential functional area is randomly selected as the destination. After the destination is determined, the simple travel chain formed is residential area-destination-residential area. Complex chain setting: First determine all residential areas. For each residential area, based on the OD matrix, first randomly select the first non-residential functional area F1 as the first intermediate destination. The selection probability is the transfer probability from the residential area to the non-residential functional area F1. Then, with F1 as the new departure point, according to the OD matrix again, a second different non-residential functional area F2 is randomly selected as the second intermediate destination. The final complex travel chain is residential area-F1-F2-residential area.

[0150] Step F: An improved Dijkstra path planning algorithm based on Pareto optimization is proposed to select paths based on user labels. The specific steps are as follows:

[0151] Step F1: Obtain the topological information of the road network based on the node data, road grade, road length and adjacency matrix of step A2, and define the path weight in multiple dimensions. Each edge corresponds to a five-tuple (t uv ,e uv ,c uv ,r uv ,s uv ), each dimension is implemented as follows:

[0152] Travel time t uv : Obtained by the improved road resistance function in step C2;

[0153] Energy consumption uv : Obtained from the improved energy consumption model of step D1;

[0154] Carbon emissions uv : Based on the mapping relationship between vehicle speed, acceleration and energy consumption, it is calculated by the formula:

[0155]

[0156] Among them, κ is the carbon emission coefficient per unit energy consumption, v avg is the average speed of the road section, λ is the speed sensitivity factor;

[0157] Road comfort uv : Build an evaluation system based on GIS data and user feedback, and convert qualitative feedback into quantitative scores;

[0158] Traffic congestion index uv : Obtain traffic congestion coefficients in different scenarios through Baidu Maps.

[0159] Step F2: Based on the multi-dimensional path weights defined in step F1, the nodes to be explored must be managed through a priority queue to ensure that each expanded path is the current comprehensive optimal solution. A priority queue is established, and the element format of the priority queue is: (node ​​number, comprehensive score F(v), predecessor node), where the predecessor node is used to record the source of the path to facilitate backtracking the complete path. Before the path begins, the starting point is added to the priority queue. The comprehensive score formula is:

[0160] F(v)=ω t ·t uv +ω e ·e uv +ω c c uv +ω r ·(1-r uv)+ω s ·s uv

[0161] Among them, (ω t ,ω e ,ω c ,ω r ,ω s ) is the weight after dynamic adjustment of reinforcement learning.

[0162] Select the node u with the smallest F(v) from the priority queue for expansion, and remove u from the priority queue. Obtain all adjacent edges (u, v) of node u based on the adjacency matrix. For each edge (u, v), calculate the new comprehensive score F′(v) of node v:

[0163] F′(v)=F(u)+ω t ·t uv +ω e ·e uv +ω c c uv +ω r ·(1-r uv )+ω s ·s uv

[0164] If v is not in the priority queue, add it to the queue and record its comprehensive score F′(v). If v is in the priority queue and F′(v) < F(v), update F(v) = F′(v) and adjust the queue order to ensure that the queue is always arranged in ascending order of comprehensive score. This process repeats until the priority queue is empty. At this point, all scalable nodes and paths have been processed. Finally, trace back from the endpoint through the predecessor nodes to obtain a set of candidate paths. Each path has a unique comprehensive score, providing the data foundation for subsequent screening.

[0165] Step F3: During the priority queue expansion process in step F2, a set of candidate paths is generated. For example, two candidate paths have solution indices of (T2, E2, C2, R2, S2) and (T1, E1, C1, R1, S1), respectively. The indices of the two solutions are determined according to the following rules: If T1 ≤ T2, E1 ≤ E2, C1 ≤ C2, R1 ≥ R2, S1 ≤ S2, and at least one inequality strictly holds (strictly greater than or strictly less than), then solution 1 Pareto dominates solution 2, solution 2 is discarded, and solution 1 is directly output as the result of the improved Dijkstra path planning algorithm. If T2 ≤ T1, E2 ≤ E1, C2 ≤ C1, R2 ≥ R1, S2 ≤ S1, and at least one inequality strictly holds, then solution 2 Pareto dominates solution 1, solution 1 is discarded, and solution 2 is directly output as the result of the improved Dijkstra path planning algorithm. If neither of these conditions is met, the two solutions do not dominate each other, and the process proceeds to step F4.

[0166] Step F4: Introduce the Q-learning reinforcement learning algorithm to achieve the target weight (ω t ,ω e ,ω c ,ω r ,ω s ), state S: defined as the combination vector of the current travel scenario and the user's historical preferences, action A: the operation of adjusting the weights of each target, and reward R: calculated based on the user's feedback on the route selection and the matching degree of the real-time traffic conditions. The weight update formula is:

[0167] ω i (t+1)=ω i (t)+η[R(t)+βmax a Q(S(t+1),A(t+1))-Q(S(t),A(t))]

[0168] Among them, η is the learning rate, β is the discount factor, and Q(S,A) is the state-action value function.

[0169] Through continuous iterative learning, the algorithm automatically adapts to user preferences and scenario changes, updates the weights in the comprehensive score formula, and then affects the node expansion priority of the Dijkstra algorithm, and restarts from step F2. The specific process is shown in the attached Figure 7 shown.

[0170] Step G: Improve user demand judgment to ensure that the remaining SOC can reach a nearby charging station; design a charging station selection model that considers comprehensive costs; the specific steps are as follows:

[0171] Step G1: The traditional charging demand determination method usually assumes that charging is required only when the current power is not enough to drive to the next destination. Improve user demand judgment to ensure that the remaining SOC Can reach a nearby charging station. The specific formula is as follows:

[0172]

[0173] Among them, S OC is the current power ratio, S dest S is the proportion of power consumption when traveling to the next destination. safe is the safe power threshold (set to 0.15), P cha is the charging probability. S dest The formula is:

[0174]

[0175] Among them, E a'b' is the energy consumption of traveling from the current location a' to the next destination b'.

[0176] When the current power ratio is less than the safe power threshold plus the power ratio corresponding to the estimated power consumption to the next destination (S OC <S safe +S dest ), it means that the vehicle’s power is not enough to reach the next destination safely. At this time, the charging probability P cha If it is judged as 1, it is considered that there is a need for charging. When the current power ratio is less than 20%, it means that the vehicle power is low and there may be a risk of insufficient power when going to the charging station, so it is also judged that there is a need for charging. cha Also 1.

[0177] Step G2: Design a charging station selection model that considers comprehensive costs. Combining traditional objective costs with psychological costs, users select the charging station with the lowest comprehensive cost. Traditional objective costs include time costs, queuing costs, charging costs, and energy loss costs. Psychological costs include familiarity effect costs, quick gratification effect costs, and loss aversion effect costs.

[0178] The current electricity proportion, queuing time, and electricity price are fuzzified. The current electricity proportion is defined as a fuzzy set of {tight, moderate, sufficient}. The membership functions of tight, moderate, and sufficient are:

[0179]

[0180] The fuzzy set of queue time is defined as {short, medium, long}, and the membership functions of short, medium, and long are:

[0181]

[0182]

[0183] The electricity price p is defined as a fuzzy set of {low, medium, high}, and the membership functions of low, medium, and high are:

[0184]

[0185] Among them, p avg is the average electricity price in the region.

[0186] Combining expert experience with user behavior data, fuzzy rules are constructed, and the Mamdani reasoning method is used for dynamic fuzzy reasoning and weight updating. The following steps are used: based on the fuzzification function, the membership of the current electricity consumption proportion, queue time, and electricity price is calculated. For the fuzzy rule, the minimum membership of each condition is taken as the trigger strength. The fuzzification is performed using the center of gravity method to calculate ζ and determine the precise dynamic weight.

[0187] Calculate the improved road resistance function T from the user's location to the candidate charging station. The physical meaning of the improved road resistance function T is the travel time, so T can be directly used as the travel time t and converted into time cost according to the time value; obtain the electricity price p of the charging station through the official website, and calculate the user's expected charging amount E based on the current remaining power and the planned replenishment power charge , calculate the charging cost; obtain the charging station queue time t through the official website queue , calculate the queuing cost according to the queuing time cost coefficient ω2; according to the improved energy consumption model in step D, the energy consumption ΔE of driving to the charging station is obtained loss , and converted into energy consumption cost according to the energy consumption cost coefficient ω3; if the charging station is known, then C f = 0, if it is a new user or a charging station that has never been used, the basic cost λ is calculated based on the familiarity f and information completeness index I info Calculate C f ; Compare the actual charging time t charge With reference charging time t ref , substitute into the formula to solve C s ; According to the real-time queue number n q and the queuing psychological cost coefficient λ q Calculate the loss aversion effect C q The formula involved is:

[0188] C t' =ζ·C o +(1-ζ)·C p

[0189] C o =C time +C cost +C queue +C loss

[0190] C time =ω1·t

[0191] C cost =p·E charge

[0192] C queue =ω2·t queue

[0193] C loss =ω3·ΔE loss

[0194] C p =C f +C s +C q

[0195]

[0196] Among them, C t' is the total decision cost, C o is the objective cost, C p is the psychological cost, C time is the time cost, C cost is the charging cost, C queue is the queuing cost, C loss is the energy cost, C f is the familiarity effect, C s For the quick gratification effect, C q is the loss aversion effect, ω1 is the user's time cost coefficient, p is the electricity price of the charging station, E charge is the estimated charge capacity, t queue is the queuing time at the charging station, ω2 is the queuing time cost coefficient, ΔE loss To calculate the energy consumption of driving to the charging station, ω3 is the energy cost coefficient, λ f is the basic cost of familiarity, I info is the information completeness index, t charge is the actual charging time, t ref is the reference charging time, λ ins is the psychological cost coefficient of instant satisfaction, n q is the real-time queue number, λ q is the psychological cost coefficient of queuing. The specific process is as shown in the attached Figure 8 shown.

[0197] Step H: Design a charging mode decision system with dual constraints of time and economy; use Monte Carlo simulation to effectively predict the regional electric vehicle charging load; the specific steps are as follows:

[0198] Step H1: The definition of available charging time must follow the time allocation principle of actual charging scenarios. The user's stay time at the destination cannot be fully used for charging, which often includes waiting time in queues. The remaining time after deducting the waiting time is the effective time that can actually be used for charging. Based on this principle, the available charging time t available The formula is:

[0199] t available =t stay -t queue

[0200] Among them, t stay The dwell time at the end point is obtained by selecting the Gaussian fitting curve corresponding to the user data set in step B2 according to the user's driving characteristics (aggressive driving or gentle driving).

[0201] Considering the rated power of the charging equipment and the actual time available for charging, the theoretical charging power P is constructed. EV , introduce the efficiency coefficient of 0.9 to reflect the loss factors in actual charging, and then combine it with the charging time to determine the theoretical charging power P EV The specific formula is:

[0202] P EV =P km ×0.9×t available

[0203] Among them, P km For charging power, fast charging is 60KW and slow charging is 7KW.

[0204] Target power ratio SOC target Limited by the upper limit of battery capacity, the specific formula is as follows:

[0205]

[0206] Among them, SOC now is the current battery SOC;

[0207] Calculate the actual charge capacity ΔE actual :

[0208] ΔE actual =C t ×(SOC target -SOC now )

[0209] Calculate the charging cost C charge :

[0210] C charge =ΔE actual ×p

[0211] Determine the target power ratio SOC of fast charging and slow charging target Is it greater than SOC? min (set to 0.7), if it is greater than, then choose the charging cost C of fast charging and slow charging charge If the lower option is less than , then select the target power ratio SOC of fast charging and slow charging. target The higher option.

[0212] Step H2: Use Monte Carlo simulation and set the number of simulations to 50,000. For each simulation, randomly select a user from the data set filtered and cleaned in step B1 to obtain the user's driving style, time priority, and energy priority labels. Based on the user's driving style, obtain the actual travel characteristic data (trip start time, destination dwell time) from the Gaussian fitting curve of the corresponding style in step B2. Randomly generate a scenario according to step C1 to obtain the road traffic in the specific scenario, and obtain the OD matrix in the specific scenario according to step E. According to step E, the randomly generated travel chain is a simple chain or a complex chain. Take the simple chain as an example: residential area-destination-residential area. According to step A1, randomly generate the user's initial node in the residential area. According to the OD matrix in step E, obtain the destination functional area, and according to step A1, randomly generate a node in the destination functional area as the destination node. Obtain the optimal route according to the improved Dijkstra path planning algorithm in step F. Step G1 determines whether the user can reach the destination node using the optimal route. If so, the user drives to the destination node and updates the current time. If not, step G2 selects a charging station based on the overall cost. The improved Dijkstra path planning algorithm in step F is used to obtain the optimal route to the charging station. Upon arrival at the charging station, step H selects a charging method, updates the current time, and the charging station records the charging load data. After charging is complete, the improved Dijkstra path planning algorithm in step F is used to obtain the optimal route to the destination node. The user drives to the destination node and updates the current time.

[0213] According to step A1, the user's destination node is randomly generated in the residential area. The improved Dijkstra path planning algorithm in step F is used to obtain the optimal route. Step G1 determines whether the user can reach the destination node using the optimal route. If so, the user drives to the destination node, updates the current time, and the trip ends. If not, a charging station is selected based on the overall cost according to step G2. The improved Dijkstra path planning algorithm in step F is used to obtain the optimal route to the charging station. Upon arrival at the charging station, the charging method is selected in step H, the current time is updated, and the charging station records the charging load data. After charging is complete, the improved Dijkstra path planning algorithm in step F is used to obtain the optimal route to the destination node. The user drives to the destination node and updates the current time. At the end of the trip, the number of simulations is incremented by 1. The Monte Carlo simulation is repeated until 50,000 times, generating a valid prediction result for the regional electric vehicle charging load, which paves the way for the subsequent sizing and site selection of electric vehicle charging stations.

[0214] The embodiments described are preferred implementations of the present invention, but the present invention is not limited to the above implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention are within the scope of protection of the present invention.

Claims

1. An electric vehicle charging load forecasting method based on multi-source spatiotemporal data and user behavior, characterized by: Import a map of a certain area into QGIS software, select the prediction area on the map, determine the road network topology within the prediction area, and divide the selected prediction area into functional areas according to POI data; Obtain user feature data, perform adaptive clustering algorithm analysis on user features based on driving style, time preference, and energy efficiency preference, and label the user; Solve the OD matrix under different scenarios and set up travel chains; In a certain scenario, based on the travel chain, the user's initial node is randomly generated in a certain functional area, the destination functional area is determined according to the OD matrix in the scenario, the destination node is randomly generated in the destination functional area, and the improved Dijkstra path planning algorithm based on Pareto optimization is used to select a path based on the user's label; when arriving at the destination node, the terminal node is randomly generated in the terminal functional area, and the improved Dijkstra path planning algorithm is used to select a path based on the user's label, and the user travels to the terminal node; In the process of heading to the next node, if there is a demand for charging, the remaining SOC of the electric vehicle determines the nearby charging stations that can be reached. A charging station selection model that considers the comprehensive cost is designed. The user selects the charging station with the lowest comprehensive cost among the nearby charging stations that can be reached. The improved Dijkstra path planning algorithm determines the optimal path to the charging station with the lowest comprehensive cost. After arriving at the charging station, the charging method is determined through a time-economy dual-constraint strategy. The charging station records the charging load data for subsequent fixed-capacity site selection of electric vehicle charging stations.

2. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 1 is characterized in that: The improved Dijkstra path planning algorithm based on Pareto optimization specifically includes: The topological information of the road network is obtained from node data, road grade, road length, and adjacency matrix, and the path weight is defined in multiple dimensions. Each edge corresponds to a five-tuple (t uv ,e uv ,c uv ,r uv ,s uv ); Establish a priority queue. The element format of the priority queue is: (node ​​number, comprehensive score F(v), predecessor node). The predecessor node is used to record the path source to facilitate backtracking the complete path. Before path planning begins, the initial node is added to the priority queue. The formula for the comprehensive score is: F(v)=ω t ·t uv +oh e ·e uv +oh c ·c uv +oh r ·(1-r uv )+ω s ·s uv Among them, (ω t ,ω e ,ω c ,ω r ,ω s ) is the weight after dynamic adjustment of reinforcement learning, t uv is the travel time, e uv is energy consumption, c uv is carbon emissions, r uv For road comfort, s uv is the traffic congestion index; Select the node u with the smallest F(v) from the priority queue for expansion, and remove u from the priority queue. Obtain all adjacent edges (u, v) of node u based on the adjacency matrix. For each edge (u, v), calculate the new comprehensive score F′(v) of node v: F′(v)=F(u)+ω t ·t uv +oh e ·e uv +oh c ·c uv +oh r ·(1-r uv )+ω s ·s uv If v is not in the priority queue, add v to the priority queue and record its comprehensive score F′(v); if v is in the priority queue and F′(v) < F(v), then update F(v) = F′(v) and adjust the priority queue order to ensure that the queue is always arranged in ascending order of comprehensive score; this process is repeated until the priority queue is empty; During the priority queue expansion process, a set of candidate paths is generated. For example, two candidate paths have solution indices of (T2, E2, C2, R2, S2) and (T1, E1, C1, R1, S1). The indices of the two solutions are determined according to the following rules: If T1≤T2, E1≤E2, C1≤C2, R1≥R2, S1≤S2, and at least one inequality strictly holds, then solution 1 Pareto dominates solution 2, solution 2 is discarded, and solution 1 is directly output as the result of the improved Dijkstra path planning algorithm; If T2≤T1, E2≤E1, C2≤C1, R2≥R1, S2≤S1, and at least one inequality strictly holds, then solution 2 Pareto dominates solution 1, solution 1 is discarded, and solution 2 is directly output as the result of the improved Dijkstra path planning algorithm; If the above two conditions are not met, the two solutions do not dominate each other, and the Q-learning reinforcement learning algorithm is introduced to achieve the target weight (ω t ,ω e ,ω c ,ω r ,ω s ) dynamic optimization. Through continuous iterative learning, the algorithm automatically adapts to user preferences and scenario changes, updates the weights in the comprehensive score formula, and then affects the node expansion priority of the Dijkstra algorithm, starting again from the establishment of the priority queue; Among them, T1, E1, C1, R1, and S1 are the driving time, energy consumption, carbon emissions, road comfort, and traffic congestion index of the first candidate path, respectively; T2, E2, C2, R2, and S2 are the driving time, energy consumption, carbon emissions, road comfort, and traffic congestion index of the second candidate path, respectively.

3. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 2 is characterized in that: The travel time t uv Obtained by the improved road resistance function, the improved road resistance function is: T=T L +T J f(T temp )=1+0.005·|T temp -20|T temp Among them, T is the improved road resistance function, T L is the section impedance, T J is the node impedance, t0 is the free travel time of the road section, y′ is the corrected saturation, Q is the road flow, C is the road capacity, λ level Road grade adjustment factor, f(T temp ) is the temperature correction function, λ habit is the driving habit coefficient, k is the signal period, τ is the green light signal ratio, y is the node saturation, λ turn is the steering influencing factor, α is the coefficient of the corrected saturation linear term, β is the coefficient of the corrected saturation quadratic term, q is the node impedance calculation parameter, and γ is the corrected saturation adjustment factor.

4. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 2 is characterized in that: The energy consumption e uv Obtained from the improved energy consumption model, the improved energy consumption model is: Among them, E base is the basic energy consumption per unit mileage of the vehicle under ideal conditions, E speed is the energy consumption term affected by speed, E ac is the additional energy consumption for the interior temperature, C B (T) is the battery capacity under the influence of temperature, C t is the actual capacity of the vehicle, δ battery is the battery temperature sensitivity coefficient, k habit is the static coefficient of the original driving habit, f aggressive is the frequency of sudden acceleration and braking obtained by the vehicle sensor, Δk is the fluctuation coefficient, P c is the cooling power of the air conditioner, P h is the air conditioning heating power, T Lab is the road impedance on road section ab, i.e., the travel time, l ab is the length of section ab, v ab is the actual vehicle speed on road section ab.

5. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 1 is characterized in that: The charging station selection model considering comprehensive costs is specifically as follows: C t' =ζ·C o +(1-ζ)·C p C o =C time +C cost +C queue +C loss C time =ω1·t C cost =p·E charge C queue =ω2·t queue C loss =ω3·ΔE loss C p =C f +C s +C q Among them, C t' is the total decision cost, C o is the objective cost, C p is the psychological cost, ζ is the weight, C time is the time cost, C cost is the charging cost, C queue is the queuing cost, C loss is the energy cost, C f is the familiarity effect, C s For the quick gratification effect, C q is the loss aversion effect, ω1 is the user's time cost coefficient, p is the electricity price of the charging station, E charge is the estimated charge capacity, t queue is the queuing time at the charging station, ω2 is the queuing time cost coefficient, ΔE loss To calculate the energy consumption of driving to the charging station, ω3 is the energy cost coefficient, λ f is the basic cost of familiarity, I info is the information completeness index, t charge is the actual charging time, t ref is the reference charging time, λ ins is the psychological cost coefficient of instant satisfaction, n q is the real-time queue number, λ q is the psychological cost coefficient of queuing.

6. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 1 is characterized in that: The remaining SOC of the electric vehicle determines the nearby charging stations it can reach, including: Among them, S OC is the current power ratio, S dest S is the proportion of power consumption when traveling to the next destination. safe is the safe power threshold, P cha is the charging probability; and E a'b' is the energy consumption of traveling from the current location a' to the next destination b', C t is the actual capacity of the vehicle; When the current power ratio is less than the safe power threshold plus the power ratio corresponding to the expected power consumption to the next destination, that is, S OC <S safe +S dest , at this time the charging probability P cha If it is judged as 1, it is considered that there is a need for charging; when the current power ratio is less than 20%, it is judged that there is a need for charging. cha Also 1.

7. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 1 is characterized in that: The strategy of determining the charging mode through the dual constraints of time and economy is specifically as follows: Considering the rated power of the charging equipment and the actual time available for charging, the theoretical charging power P is constructed. EV =P km ×0.9×t available , where P km is the charging power, t available is the rechargeable time, and t available =t stay -t queue , t stay is the terminal residence time, t queue queue times for charging stations; Limited by the upper limit of battery capacity, the target power ratio Among them, SOC now is the current battery SOC, C t is the actual capacity of the vehicle; Calculate the actual charge capacity ΔE actual :ΔE actual =C t ×(SOC target -SOC now ), calculate the charging cost C charge :C charge =ΔE actual ×p, where p is the electricity price of the charging station; Determine the target power ratio SOC of fast charging and slow charging target Is it greater than SOC? min If it is greater than, then choose the charging cost C of fast charging and slow charging. charge If the lower option is less than , then select the target power ratio SOC of fast charging and slow charging. target The higher option.

8. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 1 is characterized in that: The user characteristics are analyzed by adaptive clustering algorithm according to driving style, time preference and energy efficiency preference, and users are labeled, specifically: The original data set is obtained from traffic survey data, and relevant data of multiple brands and models commonly seen in the market are imported. Then, the data is filtered and cleaned to determine the characteristic data to distinguish driving style, time preference, and energy efficiency preference. The user's driving style is distinguished as aggressive or gentle according to mileage, acceleration from 0 to 100 km / h, and dwell time at the destination. The user's preference is distinguished as high time priority or low time priority according to the start time of the trip. The user's preference is distinguished as high energy efficiency priority or low energy efficiency priority according to the age of the vehicle, energy consumption per 100 kilometers, and fast charging power. The density peak clustering algorithm is used to automatically identify the cluster center by calculating the local density and relative distance of user feature data points, thereby achieving effective clustering of user features.

9. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 8 is characterized in that: According to the "density-distance" decision diagram, the local density ρ of all user feature data points i' and the relative distance δ i’ Sort them from large to small and select ρ i' and δ i' The two points that are both in the top 5% are used as cluster centers, and each user feature data point is assigned to the cluster center closest to it to obtain the cluster label.

10. The electric vehicle charging load prediction method based on multi-source spatiotemporal data and user behavior according to claim 1 is characterized in that: The functional areas are divided into residential area, social and entertainment area, work area, commercial area, catering, hospital, school and others.

Images