In this study I build a first-person, model-driven investigation of how electric car charging and battery-swapping demands evolve in time and space when many electric car users follow multi-layer complex travel chains through a dynamic urban road network. My central concern is practical: an electric car is not a stationary load, and its charging demand is not produced at a fixed bus at a fixed hour. Instead, an electric car carries time, space, and state-of-charge attributes as it moves, and those attributes are shaped by road congestion, functional land use, dwell time, driver urgency, and the availability of charging or swapping service. I therefore treat the electric car as the moving link between the transportation network and the distribution network. I use the phrase electric car repeatedly because the behavior of each electric car, rather than an aggregate annual average, is the elementary unit from which my forecast is constructed.
My objective is to forecast the spatiotemporal distribution of electric car charging and swapping loads under a planning horizon that represents a full day with half-hour resolution. I also aim to reveal the dynamic compensation mechanism through which battery-swapping loads can reduce the volatility of charging loads when renewable energy accommodation is considered. The model I present is deliberately modular. The first module samples the travel characteristics of private electric car users. The second module finds the optimal path through a dynamic road graph and converts traffic flow into speed, travel time, and energy consumption. The third module represents charging and swapping stations as state-dependent service alternatives. The fourth module represents wind and photovoltaic generation as uncertain but bounded sources. The fifth module aggregates the electric car loads and compares operating modes. I use tables and equations throughout so that the logic can be inspected, reproduced, and extended.

Modeling perspective and system coupling
I begin from the view that the electric car is a mobile load with three coupled attributes. The time attribute describes when the electric car departs, how long it stays at a destination, and when it returns. The space attribute describes which road nodes and functional zones the electric car visits. The state attribute describes the battery SOC and the energy required for the next trip. When the SOC of an electric car falls below a threshold, the user must choose a service action: slow charging, fast charging, or battery swapping. That choice is not independent of traffic. Congestion changes speed, speed changes energy consumption per unit distance, and energy consumption changes the SOC trajectory. At the same time, many electric car users making similar choices can concentrate load at particular stations and at particular times. I therefore couple the road graph and the power graph through the electric car state trajectory rather than through a static load profile.
I make several assumptions explicit. First, I assume that private electric car trips are generated from a set of functional zones that include residential, working, commercial, and green-buffer areas. Second, I assume that each electric car follows a multi-layer complex travel chain with one or more destinations and dwell periods. Third, I assume that the road network is represented by a directed graph with road levels, capacities, and signalized nodes. Fourth, I assume that charging and swapping stations are distributed across the network and that each station has a finite service type. Fifth, I assume that wind and solar generation are uncertain but remain inside bounded intervals. These assumptions do not remove uncertainty; instead, they make the uncertainty explicit enough for scenario-based forecasting.
The following table summarizes the principal notation I use. I keep the notation compact because the electric car state must be updated at every time step and every road segment.
| Symbol | Meaning | Role in the electric car forecast |
|---|---|---|
| R | Set of road nodes | Defines possible origins, destinations, and station access points |
| E | Set of road links | Defines directed travel between nodes |
| L | Set of road levels | Controls free-flow speed, capacity, and energy intensity |
| q_ij(t) | Traffic flow on link i-j at time t | Converts traffic state into speed and delay |
| C_ij^l | Capacity of road level l | Limits flow and influences congestion |
| v_ij(t) | Speed on link i-j at time t | Determines travel time and energy use |
| T_ij | Link travel time | Updates electric car arrival time |
| T_C | Signal node delay | Captures intersection impedance |
| d_e | Energy consumption per unit distance | Updates electric car SOC while moving |
| S_t | SOC at time t | Triggers charging or swapping decisions |
| S_min, S_max | SOC bounds | Define feasible electric car operation |
| P_t | Charging power at time t | Forms the charging load |
| P_f, P_s | Fast and slow charging power | Define station service modes |
| D_ln | Destination of layer n | Builds simple or complex travel chains |
| t_n | Dwell time at destination | Limits charging opportunity |
| u_PV(t) | Photovoltaic output uncertainty variable | Represents renewable variability |
| P_WT | Wind turbine output | Represents wind generation |
Multi-layer complex travel chains for electric car users
I describe the daily movement of a private electric car as a travel chain. A travel chain starts at an origin, visits one or more destinations, and eventually returns to a final destination. The chain is simple when the electric car moves between two functional zones in a back-and-forth pattern. The chain is complex when a functional zone can act both as an origin and as a destination in a multi-stage sequence. I use the number of destinations in the chain to distinguish the two cases. If the chain has only one destination layer, I treat it as a simple chain. If the chain has more than one destination layer, I treat it as a complex chain. This distinction matters because a complex chain creates more opportunities for the electric car to arrive with a low SOC at an inconvenient place and to require either fast charging or swapping.
I define the initial location as S_w, the initial departure time as S_t, the number of trips as T_s, the travel chain combination as L_t, the destination at level n as D_ln, the dwell time at that destination as t_n, and the road nodes traversed as R_nn. The simple and complex cases can be written as follows.
$$
\begin{cases}
\text{simple chain}, & D_{ln}=1,\\
\text{complex chain}, & D_{ln}>1.
\end{cases}
$$
For a single trip in a multi-layer chain, I track the departure time, the chosen destination, the route, the arrival time, the dwell time, and the SOC at arrival and departure. The SOC update during travel is not continuous in my implementation; it is updated segment by segment because speed and road level change by segment. This is important for an electric car because a short but congested segment can consume more energy per kilometer than a long but free-flowing segment. The following table shows how I sample the key travel-chain variables. The exact distribution parameters can be changed by scenario, but the sampling logic remains the same.
| Travel-chain variable | Sampling or assignment logic | Effect on electric car demand |
|---|---|---|
| Initial departure time | Bimodal morning and afternoon distribution | Creates morning and evening electric car load peaks |
| Initial location | Weighted by residential zone population | Determines first trip origin |
| First destination | Work, commercial, or residential according to trip purpose | Sets the first charging opportunity |
| Dwell time | Destination-specific distribution | Limits slow charging and favors fast or swap service |
| Return destination | Residential zone with high probability | Creates overnight electric car charging load |
| Number of trips | One to three layers | Controls complexity and SOC stress |
| Service choice | Urgency, SOC, dwell time, and station availability | Selects slow charge, fast charge, or swap |
I treat the travel chain as a generator of charging opportunities rather than as a fixed schedule. An electric car with a long dwell time at a residential or working destination can use slow charging. An electric car with a short dwell time but a low SOC is more likely to use fast charging. An electric car with a short dwell time, a low SOC, and an urgent next trip is the natural candidate for battery swapping. This is the behavioral foundation of the charging-swapping interaction mode that I develop later. Because every electric car is sampled separately, the aggregate forecast is not simply an average; it is the sum of many state-dependent trajectories.
Dynamic traffic network and optimal path search
I represent the urban road system as a directed graph. The graph contains nodes, links, and road levels. The road level controls free-flow speed, capacity, and the coefficients used in the energy-consumption model. I write the graph as follows.
$$
\begin{cases}
G=(R,E,L),\\
R=\{r_i \mid i=1,2,3,\ldots,n\},\\
E=\{r_{ij} \mid r_i \in R,\ r_j \in R,\ i \ne j\},\\
L=\{r_{ij}^{l} \mid l=1,2,\ r_{ij}\in E\}.
\end{cases}
$$
Here R is the set of road nodes, E is the set of road links, and L is the set of road levels. The superscript l distinguishes a level-one road from a level-two road. I use an adjacency matrix to store connectivity and link attributes. A link that does not exist is assigned an infinite cost, while an existing link stores its length, capacity, free-flow speed, and level. With this representation, the shortest path problem for an electric car becomes a weighted graph search problem. I use the Floyd algorithm because it computes all-pairs shortest paths and because it is convenient when many electric car origins and destinations are sampled repeatedly. The algorithm updates the path cost through intermediate nodes and retains the best predecessor for route reconstruction.
$$
d_{ij}^{(k)}=\min\left(d_{ij}^{(k-1)},\ d_{ik}^{(k-1)}+d_{kj}^{(k-1)}\right).
$$
The cost used in the shortest path search is not simply distance. I use a generalized travel cost that includes link travel time, signal delay, and an energy penalty. This matters for an electric car because the fastest route may not be the most energy-efficient route. In my model the route choice can be tuned between time priority and energy priority. When the electric car is in a normal SOC state, time priority dominates. When the electric car is close to its lower SOC bound, energy priority becomes stronger. This adaptive weighting is one way in which the electric car state feeds back into route choice.
Traffic flow is dynamic. I use a wavelet neural network to forecast traffic flow from historical and real-time inputs. The network input and output are normalized, and the network is trained by backpropagation. I do not treat the forecast as perfect; I treat it as the expected traffic state for the current time step. The predicted flow is then converted into speed by a flow-speed relationship. The relationship has a free-flow branch and a congestion-sensitive branch. I write the speed model as follows.
$$
\begin{cases}
\beta=a+b\left(\dfrac{q_{ij}(t)}{C_{ij}^{l}}\right)^{\eta},\\[6pt]
v_{ij}(t)=\dfrac{v_{l0}}{1+\left(\dfrac{q_{ij}(t)}{C_{ij}^{l}}\right)^{\beta}},
\end{cases}
$$
where q_ij(t) is the real-time traffic flow on link i-j, C_ij^l is the maximum capacity of a road of level l, v_l0 is the zero-flow speed of that road level, and a, b, eta, and beta are calibrated coefficients. The link travel time follows directly from the link length and the link speed.
$$
T_{ij}=\dfrac{r_{ij}}{v_{ij}(t)}.
$$
Signalized intersections add delay even when the link itself is not congested. I model the node impedance as a function of capacity, green ratio, and flow ratio. The expression I use is an average control delay form. I keep it in hours so that it can be added directly to link travel time.
$$
T_C=0.441\left[\dfrac{C_{ij}^{l}(1-\lambda)^2}{2(1-\lambda)\dfrac{q_{ij}(t)}{C_{ij}^{l}}+\dfrac{q_{ij}(t)}{C_{ij}^{l}}^2}+6.996\right],
$$
where lambda is the green split at the node. For a single trip, the total travel time is the sum of link travel times plus the sum of signal delays at intermediate nodes.
$$
T_h=\sum_{n=1}^{N}T_{ij}^{n}+T_{C}^{n,n+1}.
$$
The energy consumption of an electric car depends on speed, road level, and distance. I use a piecewise unit-mile energy model. Level-one roads and level-two roads have different coefficients because their speed profiles and stop-and-go patterns differ. The model is calibrated so that low speed and high speed both increase energy use per kilometer relative to a moderate speed range.
$$
d_e=
\begin{cases}
0.247+\dfrac{1.52}{v_{ij}}-0.004v_{ij}+2.992\times10^{-5}v_{ij}, & l=1,\\[8pt]
-0.179+0.004v_{ij}+\dfrac{5.492}{v_{ij}}, & l=2.
\end{cases}
$$
The SOC update for a moving electric car is then the previous SOC minus the energy consumed on the segment divided by battery capacity. I write this as a discrete update because the forecast is performed in half-hour steps and because each electric car may complete several segments within a step.
$$
S_{t+\Delta t}=S_t-\dfrac{d_e\,r_{ij}}{C}.
$$
The following table summarizes the traffic and energy modules. I use this table as a consistency check when I implement the electric car simulation, because each symbol in the text must have a concrete role in the code.
| Module | Input | Output | Why it matters for the electric car |
|---|---|---|---|
| Graph construction | Nodes, links, levels, lengths | Directed road graph | Defines feasible electric car movement |
| Traffic forecast | Historical and current flow | Predicted link flow | Captures time-varying congestion |
| Flow-speed conversion | Flow, capacity, road level | Link speed | Changes travel time and energy use |
| Signal delay | Green ratio, flow ratio | Node impedance | Adds realistic urban delay |
| Path search | Generalized link cost | Optimal route | Links origin, destination, and charging opportunity |
| Energy model | Speed, distance, road level | Unit-mile energy | Updates SOC and triggers service demand |
Charging and battery-swapping service models
I model two service alternatives for the electric car. The first is plug-in charging, which can be slow or fast. The second is battery swapping, which replaces a depleted battery with a charged one. A charging station and a swapping station both aim to raise the electric car SOC to a usable level, but they do so with different time scales and different grid impacts. I define the target SOC at the end of service as a fraction of the maximum SOC.
$$
S_{\text{end}}=\eta_c S_{\max}.
$$
For plug-in charging, the SOC evolves according to the charging power and the battery capacity. I also enforce the upper and lower SOC bounds and the available charging window. The charging window is important because an electric car parked for only a short time cannot complete a slow charge, even if slow charging is cheaper and gentler on the battery.
$$
\begin{cases}
S_t^{c+1}=\dfrac{S_t^{c}C+\eta_c P_t}{C},\\
S_{\min}\le S_t^{c}\le S_{\max},\\
T_{\text{begin}}^{c}\le t\le T_{\text{end}}^{c}.
\end{cases}
$$
For battery swapping, I model the swap itself as a short service event. The electric car leaves the station with a high SOC, while the station receives a depleted battery. The depleted battery then enters a charging queue. The swap time is short compared with plug-in charging, which is the main reason an electric car user may prefer swapping when the next trip is urgent.
$$
\begin{cases}
S^{r}=0.9S_{\max}^{r},\\
S_{\min}\le S_t^{r}\le S_{\max},\\
0<t\le $$="" 3\=""
The station-side battery queue is a critical part of the model. A simple strategy charges every depleted battery as fast as possible. That strategy produces many full batteries quickly, but it can create a strongly uneven SOC distribution inside the station. To avoid that, I use a two-stage queue. If a battery SOC is below a threshold, it receives fast charging. If it is above the threshold but below the target, it receives slow charging. This preserves service capability while smoothing the station load. The charging load of the station is the integral of fast and slow power over the corresponding time windows.
$$
P_{\text{swap}}(t)=
\begin{cases}
\displaystyle\int_{t_b}^{(0.2-S_t^{r})/P_f}P_s\,dt+
\int_{(0.2-S_t^{r})/P_f}^{(0.9-S_t^{r})/P_s}P_f\,dt, & 0<s_t^{r}<0.2,\\[12pt] $$="" &="" 0.2\le="" <p="" \displaystyle\int_{t_b}^{(0.9-s_t^{r})="" \end{cases}="" p_f}p_f\,dt,="" s_t^{r}I emphasize that the two integrals above are not merely a mathematical convenience. They express the physical fact that a depleted battery first absorbs high power and then absorbs lower power as its SOC rises. For an electric car user, the visible event is a short swap. For the grid, the invisible event is a longer battery-charging process that can be shifted in time. This separation between user-side service time and grid-side charging time is the source of the smoothing effect I investigate.
The following table summarizes the service modes. I use it to keep the difference between user behavior and grid behavior clear.
| Service mode | User-side duration | Grid-side load | Typical electric car use case |
|---|---|---|---|
| Slow charging | Long dwell | Low power, long duration | Residential overnight or workplace daytime |
| Fast charging | Short to medium dwell | High power, short duration | Commercial stop or urgent top-up |
| Battery swapping | Very short | Deferred battery charging | Urgent trip, limited dwell, low SOC |
Charging-swapping interaction and state transitions
I now define the interaction mode. The principle is that charging is the primary service and swapping is the supplementary service. An electric car first considers slow charging if its dwell time is long enough. If slow charging cannot deliver enough energy before departure, the electric car considers fast charging. If neither slow nor fast charging can satisfy the next trip requirement, or if the remaining dwell time is shorter than the minimum fast-charge time, the electric car considers swapping. This decision rule avoids treating swapping as a universal replacement for charging. It also reflects the practical reality that swapping stations are fewer and battery types are not fully standardized.
I formalize the decision by comparing the energy that can be delivered before departure with the energy required for the next trip. Let E_req be the required energy and E_serv be the energy that can be delivered by a candidate service within the available time. The electric car chooses the first feasible service in the order slow charge, fast charge, swap. If no service is feasible, the electric car waits or adjusts its route, and the SOC constraint becomes binding.
$$
\text{service}=
\begin{cases}
\text{slow charge}, & E_{\text{slow}}\ge E_{\text{req}},\\
\text{fast charge}, & E_{\text{slow}}<e_{\text{req}},\ $$="" &="" <p="" \end{cases}="" \text{swap},="" e_{\text{fast}}<e_{\text{req}}.="" e_{\text{fast}}\ge="" e_{\text{req}},\\="" e_{\text{slow}}The state transition of a charging station includes five events: fast charging, slow charging, SOC holding, electric car arrival, and electric car departure. The state transition of a swapping station includes four events: station-side battery charging, electric car SOC holding, electric car arrival, and electric car departure. I keep the state diagrams in mind as event logic rather than as visual decoration. When an electric car arrives with a low SOC, the station event changes the load state. When the electric car departs with a target SOC, the station event releases service capacity. The difference between arrival and departure times determines how much load is actually observed in each half-hour interval.
$$
L_{\text{station}}(t)=\sum_{k=1}^{N_{\text{car}}(t)}P_k(t)+\sum_{b=1}^{N_{\text{battery}}(t)}P_b(t).
$$
Here the first sum is the plug-in charging load of electric cars at the station, and the second sum is the battery-charging load associated with swapped batteries. This distinction is essential. If I only counted the electric car at the charger, I would miss the deferred load of the swapped battery. If I only counted the battery queue, I would miss the immediate fast-charging load of an electric car that chooses plug-in service. My forecast therefore includes both loads and then compares them under different operating modes.
Wind and photovoltaic generation under uncertainty
I include renewable generation because the value of swapping is not only to reduce waiting time for the electric car user, but also to shift charging demand toward periods with more renewable output. I model photovoltaic generation from the standard current-voltage relationship. The module current depends on short-circuit current, open-circuit voltage, maximum power point current and voltage, and two correction coefficients. The equations below define the module characteristic and the correction coefficients.
$$
\begin{cases}
I=I_{SC}-C_1I_{SC}\left[\exp\left(\dfrac{U}{C_2U_{OC}}\right)-1\right],\\[8pt]
C_1=\left(1-\dfrac{I_m}{I_{SC}}\right)\exp\left(-\dfrac{U_m}{C_2U_{OC}}\right),\\[8pt]
C_2=\left(\dfrac{U_m}{U_{OC}}-1\right)\left[\ln\left(1-\dfrac{I_m}{I_{SC}}\right)\right]^{-1}.
\end{cases}
$$
I represent photovoltaic uncertainty by an interval around the expected output. The interval is bounded by the maximum positive and negative deviation. This allows me to test whether the electric car charging-swapping schedule remains acceptable when solar output is lower than expected.
$$
u_{PV}(t)\in\left[U(t)-\Delta u_{PV}^{\max}(t),\ U(t)+\Delta u_{PV}^{\max}(t)\right].
$$
For wind generation, I use a Weibull distribution for wind speed. The shape parameter and scale parameter determine the probability of low, medium, and high wind speeds. The wind turbine output is then piecewise linear between cut-in speed and rated speed, constant between rated speed and cut-out speed, and zero outside the operating range.
$$
f(v_w)=\dfrac{k}{c}\left(\dfrac{v_w}{c}\right)^{k-1}e^{-(v_w/c)^k}.
$$
$$
P_{WT}=
\begin{cases}
0, & v_w\le v_{ci}\ \text{or}\ v_w>v_{co},\\[6pt]
P_r\dfrac{v_w-v_{ci}}{v_r-v_{ci}}, & v_{ci}<v_w\le $$="" &="" <p="" \end{cases}="" p_r,="" v_r,\\[8pt]="" v_rThe following table lists the renewable parameters I use in the scenario analysis. I do not claim that these values are universal; I use them to create a transparent test case in which the electric car load and the renewable output can be compared on the same time axis.
| Parameter | Value used in the test case | Interpretation |
|---|---|---|
| Photovoltaic rated capacity | Scenario dependent | Scales solar output |
| Wind turbine rated power | Scenario dependent | Scales wind output |
| Cut-in wind speed | 3 m/s | Below this speed, wind output is zero |
| Rated wind speed | 12 m/s | Output reaches rated power |
| Cut-out wind speed | 25 m/s | Above this speed, turbine stops |
| Wind shape parameter | 2.0 | Controls distribution shape |
| Wind scale parameter | 8.0 m/s | Controls typical wind speed |
| PV uncertainty bound | Plus or minus a percentage of expected output | Tests robustness of electric car load |
Forecast procedure and computational logic
I implement the forecast as a nested simulation. The outer loop samples electric car users. The middle loop steps through the travel chain. The inner loop traverses road segments, updates SOC, checks service thresholds, and records station load. I do not aggregate before the station-level load is formed, because aggregation too early would erase the spatial differences that matter for an electric car. I record load by station, by functional zone, and by time step. I then sum the recorded loads to obtain the system-level profile.
The following table presents the procedure in a compact form. I use it to make the electric car forecast auditable and repeatable.
| Step | Operation | Electric car state updated | Output recorded |
|---|---|---|---|
| 1 | Sample initial location, departure time, and chain length | Origin, time, trip count | Individual travel plan |
| 2 | Select destination and dwell time | Destination layer, parking duration | Charging opportunity |
| 3 | Run Floyd search with generalized cost | Route and arrival time | Link sequence |
| 4 | Traverse links and update SOC | Speed, distance, energy use | Segment energy consumption |
| 5 | Check SOC and dwell time | Service feasibility | Slow, fast, or swap choice |
| 6 | Apply station queue and battery charging | Station load state | Charging and swapping load |
| 7 | Advance time step and repeat | Next trip or return | Half-hour load profile |
| 8 | Aggregate by zone and system | No state change | Spatiotemporal load map |
Test case setup
I use an urban area with a road network of 64 nodes and 98 links. The road hierarchy contains level-one and level-two roads. The area is long in the east-west direction and narrower in the north-south direction, which creates corridors of high traffic flow and corridors of lower traffic flow. I divide the area into residential, working, commercial, and green-buffer zones. Charging and swapping stations are placed at grid nodes so that every electric car has a possible service point, but the service capacity and the functional context differ by node. The total vehicle population is 70,000, and the electric car penetration is 10 percent. I simulate a full day with half-hour resolution.
The charging and swapping parameters are listed in the table below. I keep the slow and fast charging powers the same for charging stations and swapping-station battery queues so that the comparison between modes is not distorted by arbitrary differences in power ratings.
| Service type | Charging category | Power | Typical role |
|---|---|---|---|
| Charging station | Slow charging | 3.7 kW | Long-dwell electric car charging |
| Charging station | Fast charging | 30.0 kW | Short-dwell electric car top-up |
| Swapping station | Slow battery charging | 3.7 kW | Station-side queue above threshold |
| Swapping station | Fast battery charging | 30.0 kW | Station-side queue below threshold |
I calibrate the road parameters so that level-one roads have higher free-flow speed and higher capacity than level-two roads. The coefficients in the flow-speed model and the energy model are kept consistent with this hierarchy. The following table gives the road-level values used in the test case.
| Road level | a | b | eta | Free-flow speed | Capacity |
|---|---|---|---|---|---|
| Level one | 1.726 | 3.15 | 3 | 60 km/h | 1560 PCU |
| Level two | 2.076 | 2.87 | 3 | 40 km/h | 1450 PCU |
I define three operating scenarios for comparison. The first scenario is charging-only. The second scenario is the charging-swapping interaction mode with a priority for charging. The third scenario is the charging-swapping interaction mode with renewable generation explicitly included. The scenarios are summarized below.
| Scenario | Plug-in charging | Battery swapping | Renewable generation | Purpose |
|---|---|---|---|---|
| S1 | Included | Not included | Not included | Baseline electric car load |
| S2 | Included | Included as supplement | Not included | Measure the effect of charging-swapping interaction |
| S3 | Included | Included as supplement | Included | Measure renewable accommodation and smoothing |
Results: temporal and spatial electric car load behavior
I find that the electric car charging load has clear temporal peaks and clear spatial concentration. The system-level load shows a three-peak pattern. The first peak appears late at night or very early in the morning, the second peak appears around midday, and the third peak appears in the evening. This pattern is not an arbitrary curve I impose; it emerges from the sampled departure times, the dwell-time distributions, and the service choices of individual electric cars. The late-night peak is associated with residential charging after the last trip of the day. The midday peak is associated with commercial and workplace charging during lunch or errands. The evening peak is associated with returning electric cars and pre-trip preparation for the next day.
The spatial distribution is more uneven than the temporal distribution. Some nodes serve residential zones and therefore absorb a large amount of slow charging during long dwell periods. Other nodes serve commercial zones and absorb more fast charging during short stops. Working zones sit between these two extremes. The electric car load at a node is therefore not explained by the number of electric cars alone; it is explained by the functional role of the node and by the dwell-time profile it creates. I observe that the east-west corridor nodes carry a high share of the electric car load because they connect multiple functional zones and because their road capacity permits higher traffic flow. This does not mean that every electric car on that corridor charges there; it means that the corridor creates more arrival events at nearby stations.
The following table summarizes the daily electric car charging energy by functional zone in the baseline charging-only scenario. The numbers are not intended to be universal; they show the relative order of magnitude that my model produces.
| Functional zone | Daily electric car charging energy | Dominant service pattern | Main reason |
|---|---|---|---|
| Residential zone | 460.6 MWh | Slow charging | Large population, long dwell, overnight electric car charging |
| Working zone | 217.8 MWh | Slow and fast charging | Fixed working hours, medium dwell, commuting electric car flow |
| Commercial zone | 190.4 MWh | Fast charging | Short dwell, high turnover, urgent electric car top-up |
The residential zone dominates the daily energy because the electric car stays there for a long time and can use slow charging. The working zone is second because the electric car stays for several hours and can charge during the workday. The commercial zone is third in total energy even though its instantaneous fast-charging load can be high, because the dwell time is short and the number of long charging events is limited. I emphasize that the commercial zone is not unimportant. Its fast-charging events are concentrated, and concentration is exactly what creates local peaks and power-quality concerns.
When I compare the electric car charging profile with the base load of each functional zone, I find that the interaction differs by zone. In the residential zone, the electric car charging peak overlaps with the residential base-load peak in the late afternoon and evening. This overlap increases the peak-to-valley difference and stresses the distribution feeder. In the commercial zone, the electric car load is smaller but more concentrated in the late afternoon and evening. In the working zone, the electric car load follows the working schedule and peaks around the middle of the day. These differences confirm that a single system-level forecast is not enough for planning. The same amount of electric car energy can have very different grid consequences depending on where and when it is delivered.
Results: charging-swapping interaction and smoothing
I now turn to the interaction mode. When swapping is available as a supplement, the electric car load changes in two ways. First, some fast-charging events are replaced by short swapping events. Second, the swapped batteries create a deferred battery-charging load at the station. The user-side service time becomes very short, but the grid-side load does not disappear. Instead, it is shifted and reshaped. This is the dynamic compensation mechanism I set out to examine.
I find that the swapping load is generally lower in magnitude than the plug-in charging load, but it is more evenly distributed in time. The swapping station load is concentrated at a few nodes because swapping stations are fewer than charging stations, but the battery queue smooths the load at those nodes. The fast-charging part of the battery queue handles the very low SOC batteries, while the slow-charging part handles the batteries that are already above the threshold. As a result, the station does not draw maximum power for all depleted batteries at once. The following table compares the two modes in terms of peak and variation indicators. The indicators are normalized so that the comparison is not dominated by absolute scale.
| Indicator | Charging-only mode | Charging-swapping interaction mode | Interpretation |
|---|---|---|---|
| System load peak | Higher | Lower | Swapping shifts some load away from the plug-in peak |
| Peak-to-valley difference | Higher | Lower | Deferred battery charging fills some valleys |
| Load variance | Higher | Lower | Interaction mode is smoother |
| Fast-charge concentration | Higher | Lower | Urgent electric car events move to swapping |
| Renewable overlap in midday | Limited | Improved | Battery charging can absorb more solar output |
The spatial effect is also visible. In the interaction mode, the commercial-zone load decreases because some short-dwell electric cars choose swapping instead of fast charging. The residential-zone load becomes smoother because deferred battery charging and overnight slow charging fill part of the valley. The working-zone load remains relatively stable because the working-zone dwell time is already long enough for charging. This is an important finding: swapping does not simply move load from one place to another at random. It moves load according to the functional context. Commercial zones generate urgent short-dwell electric car demand, while residential zones provide long-dwell windows that can absorb deferred battery charging.
I also examine the station-side battery queue. If all depleted batteries are charged at maximum power, the station load spikes immediately after a cluster of swaps. If the queue is split by SOC threshold, the station load is lower and longer. The electric car user still receives a full battery quickly, but the grid sees a smoother load. This is a rare case in which the user-side quality of service and the grid-side smoothing objective can be aligned, provided that the station has enough batteries and enough charger capacity. I regard this alignment as a central practical advantage of the charging-swapping interaction mode.
Results: renewable accommodation and mode comparison
When I include wind and photovoltaic generation, the comparison changes. The renewable output is not constant. Photovoltaic output rises toward midday and falls in the evening, while wind output can be more distributed across the day. The electric car charging-only load tends to peak in the evening, when photovoltaic output is falling. The interaction mode moves some battery charging into the midday and early afternoon window, where solar output is stronger. This improves the overlap between electric car load and renewable generation.
I compare three profiles: the charging-only electric car load, the charging-swapping interaction load, and the net load after renewable generation. The net load is not simply the electric car load minus renewable output, because the base load also consumes part of the renewable generation. I therefore treat the electric car load as one component of the total load and examine whether the interaction mode reduces the residual load variation. The following table summarizes the qualitative result.
| Comparison item | Charging-only mode | Charging-swapping mode with renewables | Reason |
|---|---|---|---|
| Midday electric car load | Moderate | Higher but better matched to solar | Deferred battery charging absorbs midday output |
| Evening electric car load | High | Lower | Some urgent events are served by swapping earlier |
| Residual load peak | Higher | Lower | Load shift reduces the evening ramp |
| Load smoothing | Limited | Improved | Battery queue provides temporal flexibility |
| Renewable curtailment risk | Higher | Lower | Electric car charging is shifted toward high-output periods |
I find that the smoothing effect is not unlimited. It depends on the number of batteries in the swapping station, the charger power, the service threshold, and the driving demand for swaps. If the station has too few spare batteries, the queue cannot shift load. If the station has too many spare batteries, the capital cost becomes dominant. The practical design question is therefore not whether swapping is always better, but how much swapping capacity and battery inventory are needed to obtain a given reduction in peak and variance for a given electric car population. My model can be used to explore that trade-off because it records both the user-side service events and the grid-side battery-charging events.
Interpretation of the electric car results
The first interpretation is that electric car load is intrinsically spatiotemporal. A forecast that reports only a daily total cannot tell a distribution planner where the electric car load will appear. A forecast that reports only an average spatial distribution cannot tell an operator when the electric car load will peak. My model therefore reports load by station, by functional zone, and by time step. This granularity is necessary because the electric car is a moving object and because the transportation network redistributes demand across space.
The second interpretation is that the functional zone is a strong predictor of electric car service choice. Residential zones favor slow charging because the electric car stays for a long time. Commercial zones favor fast charging or swapping because the electric car stays for a short time. Working zones favor a mixture because the dwell time is intermediate. This is why I do not use a single charging probability for all electric cars. I let the travel chain and the dwell time determine the feasible service set, and I let the SOC and urgency determine the final choice.
The third interpretation is that battery swapping is a compensating service rather than a universal substitute. It compensates for limited dwell time and for urgent electric car demand. It also compensates for the temporal mismatch between renewable output and charging demand. The compensation is not free. It requires battery inventory, charger capacity, and a queue-management policy. My model shows that a two-stage queue based on SOC threshold can provide smoother station load than a pure fast-charging queue. This is a useful design insight for electric car service networks.
The fourth interpretation is that renewable generation changes the value of flexibility. In a system with low renewable penetration, the main objective may be to reduce the electric car peak. In a system with high renewable penetration, the main objective may be to move electric car load toward periods of high renewable output. The charging-swapping interaction mode can serve both objectives to some extent because the swapped battery is a buffer. The electric car user does not need to wait for the battery to charge, while the grid can choose when to charge the battery. This decoupling of user time and grid time is the most important property of the interaction mode.
Sensitivity and design implications
I perform a qualitative sensitivity analysis by changing the key parameters one at a time. The results are summarized below. I use qualitative labels rather than exact numbers because the purpose is to identify which parameters deserve detailed calibration in a real electric car deployment.
| Parameter change | Effect on electric car charging load | Effect on swapping load | Design implication |
|---|---|---|---|
| Higher electric car penetration | Higher peak and higher energy | More swap events | Station capacity must scale with electric car adoption |
| Longer residential dwell | More slow charging | Fewer urgent swaps | Residential charging can absorb overnight load |
| Shorter commercial dwell | More fast charging | More swap demand | Commercial zones need fast and swap options |
| Higher traffic congestion | Higher energy use per mile | More swap demand | Traffic management indirectly affects electric car load |
| Lower SOC threshold | More urgent charging | More swaps | Driver risk behavior matters |
| Larger battery inventory | Lower immediate charging peak | More deferred charging | Inventory is a flexibility resource |
| Higher fast-charge power | Higher local peaks | Shorter battery queue | Power quality must be checked locally |
| More midday solar output | Better overlap if load is shifted | More useful battery charging | Swapping can improve renewable accommodation |
I draw several design implications from this sensitivity analysis. First, the electric car charging network should not be planned by total energy alone. The location of slow charging matters because slow charging is sensitive to dwell time. The location of fast charging matters because fast charging creates local peaks. The location of swapping matters because swapping creates deferred battery load. Second, the station queue policy is as important as the station power rating. A station with a well-designed SOC threshold queue can provide a smoother load than a station with a higher power rating but no queue control. Third, the transportation network and the power network must be planned together. A congested road reduces speed, increases energy use per mile, and increases the probability that an electric car will need urgent service. That urgent service can appear as a fast-charging peak or as a swap cluster, depending on the station mix.
Limitations and next steps
I recognize several limitations. The first limitation is that I do not optimize the placement of charging and swapping stations. I place stations at grid nodes so that the electric car can access service, but I do not solve a formal facility-location problem. This means that some node-level differences in my results reflect the initial station layout rather than the true optimal layout. In future work I would couple the forecast with a siting and sizing model so that station placement responds to the electric car load distribution.
The second limitation is that I use a bounded interval for renewable uncertainty rather than a full stochastic optimization. The interval is useful for testing robustness, but it does not produce a probability distribution of net load. A two-stage or multi-stage stochastic program could be added to determine the best charging schedule for the battery queue under uncertain wind and solar output. I would keep the electric car travel-chain sampler as the scenario generator and use the stochastic program to control the station-side charging decisions.
The third limitation is that I do not model driver response to price or to queue length in full detail. An electric car user may choose a different station if the queue is long, or may postpone charging if the price is high. I use a feasibility-based service choice rather than a full utility model. A next step would be to add a discrete-choice model in which the electric car user weighs time, price, convenience, and risk. That would make the interaction between charging and swapping more behaviorally realistic.
The fourth limitation is that I do not include vehicle-to-grid discharge. A swapping station with a large battery inventory could, in principle, support the grid during peak periods. I do not model that reverse flow here. I note it as an extension because the same battery inventory that smooths charging could also provide short-term peak shaving. The electric car would remain the primary service object, but the station battery would become a grid asset as well.
Concluding synthesis
I conclude that the spatiotemporal forecast of electric car charging and swapping loads is best treated as a coupled transportation-power problem. The electric car is the unit that connects the road network to the distribution network. Its SOC trajectory is shaped by traffic flow, road level, signal delay, dwell time, and service availability. A forecast that ignores any of these factors will miss part of the load. By using a multi-layer complex travel chain, I capture the sequence of destinations and dwell periods that produces electric car charging opportunities. By using the Floyd algorithm on a dynamic graph, I capture the route choice and the travel time that determine when the electric car arrives at a station. By using a flow-speed model and a piecewise energy model, I capture the effect of congestion on SOC. By using a charging-swapping interaction mode, I capture the difference between user-side service time and grid-side battery charging time.
The main empirical finding is that electric car charging demand has strong regional differences. Residential zones produce large slow-charging energy, working zones produce intermediate load, and commercial zones produce concentrated fast-charging or swapping demand. These differences are not accidental. They follow from the functional role of each zone and from the dwell-time profile that each zone imposes on the electric car. A station that serves a commercial zone faces a different load shape from a station that serves a residential zone, even if both stations have the same charger rating.
The second main finding is that battery swapping can effectively mitigate charging-load fluctuations when it is used as a supplement rather than as a universal replacement. The swapped battery creates a deferred load that can be shifted toward periods of higher renewable output. The station-side SOC threshold queue prevents all depleted batteries from drawing maximum power at once. The result is a smoother load, a lower peak-to-valley difference, and better overlap with midday solar generation. The electric car user benefits from short service time, and the grid benefits from temporal flexibility.
The third main finding is that renewable generation changes the design objective. When renewable penetration is high, the question is not only how to reduce the electric car peak but also how to move electric car load toward high-output periods. The charging-swapping interaction mode provides a practical mechanism for doing so because the battery is an intermediate buffer. The electric car does not need to wait for the battery to charge, while the station can choose when to charge it. This decoupling is the core of the dynamic compensation mechanism I set out to examine.
I therefore recommend that electric car infrastructure planning consider three layers together: the travel-chain layer, the traffic-network layer, and the station-service layer. The travel-chain layer determines who needs service and when. The traffic-network layer determines how much energy the electric car consumes before it reaches service. The station-service layer determines whether the electric car uses slow charging, fast charging, or swapping. When these three layers are coupled, the resulting forecast is more informative than a static load curve. It shows where the electric car load will appear, when it will appear, and how it can be reshaped by service design and renewable supply.
Finally, I note that the electric car is not only a load. It is also a source of flexibility. A charging station can shift some of its demand through scheduling. A swapping station can shift more demand through battery inventory. If future stations are allowed to discharge to the grid, the electric car and its battery can become an even more active participant in distribution-system operation. My model provides a foundation for that extension because it tracks the electric car SOC, the station battery SOC, and the time at which each service event occurs. With those state variables, a future controller can decide not only when to charge but also whether to hold, discharge, or reserve capacity for the next electric car arrival.
</s_t^{r}
