Spatiotemporal EV Charge-Swap Load Prediction

I develop a first-person, model-driven framework for predicting the spatiotemporal distribution of electric vehicle charging and battery-swapping loads in an urban distribution network with renewable generation. The central motivation is straightforward: as electric vehicle penetration rises, uncontrolled charging can amplify peak demand, increase reverse power flows, and reduce the flexibility that renewable energy sources could otherwise provide. I therefore treat the electric vehicle not only as a mobile load but also as a moving agent whose travel purpose, route choice, state of charge, dwell time, and decision between charging and swapping jointly determine when and where load appears. The framework links a multi-layer complex travel chain, a dynamic road network, a charge-swap interaction mode, and wind-photovoltaic generation models. I use the model to answer three questions: how electric vehicle charging demand differs across functional areas, how battery swapping can compensate for charging-load fluctuations, and how the combined charge-swap pattern interacts with renewable output.

My modeling logic begins with individual electric vehicle trips and ends with aggregated nodal loads. I sample initial locations, initial travel times, initial state of charge, trip counts, destination sequences, and dwell durations. I then propagate each electric vehicle through a road graph whose edge travel times depend on real-time traffic flow. When the state of charge reaches a decision threshold, the electric vehicle either continues, charges slowly, charges quickly, or swaps its battery, depending on urgency, remaining dwell time, and the expected energy required for the next trip. The load of each station is accumulated over time and space. Finally, I compare the charge-only pattern with the charge-swap pattern and evaluate both against wind and photovoltaic output.

The main contributions I claim are as follows. First, I construct a multi-layer complex travel chain that represents repeated trips among residential, workplace, commercial, and green areas rather than a single origin-destination pair. Second, I couple the Floyd shortest-path algorithm with a flow-speed relation and a node-impedance model, so that electric vehicle energy consumption reflects traffic congestion and road class. Third, I formulate a charge-swap interaction mode in which fast charging, slow charging, and battery swapping are selected endogenously. Fourth, I model the deferred charging of depleted swap batteries through a threshold-based queue, which allows the swapping station to shift part of its load away from high-price or high-net-load periods. Fifth, I evaluate the resulting electric vehicle load against renewable generation and show why the swap mode is useful for valley filling and peak shaving.

I use a consistent notation throughout. The road graph is represented by a set of nodes, a set of edges, and a set of road classes. The travel chain is represented by a sequence of destinations and dwell times. The electric vehicle state is represented by its state of charge, remaining dwell time, and next-trip energy requirement. The station state is represented by the number of connected electric vehicles, the number of available batteries, and the state of charge of each battery in the charging queue. The renewable system is represented by photovoltaic irradiance, ambient temperature, wind speed, and the resulting available power. The following table summarizes the main symbols I use in the model.

Main symbols used in the electric vehicle charge-swap prediction model
Symbol Meaning Typical unit
G Road network graph dimensionless
R Set of road nodes dimensionless
E Set of road edges dimensionless
L Set of road classes dimensionless
r_ij Road edge between node i and node j km
q_ij(t) Traffic flow on edge r_ij at time t PCU
C_ij^l Capacity of a class-l road edge PCU
v_ij(t) Average speed on edge r_ij at time t km/h
T_ij Travel time on edge r_ij h
T_C Node impedance at a signalized intersection h
d_e Unit-distance energy consumption kWh/km
S_t State of charge at time t p.u.
P_t Charging power at time t kW
P_f Fast charging power kW
P_s Slow charging power kW
T_h Total impedance of one trip h
N Number of road segments in one trip dimensionless
u_PV Photovoltaic output uncertainty variable kW
P_WT Wind turbine output kW
v_w Wind speed m/s
k, c Weibull shape and scale parameters dimensionless, m/s

I define the road network as a directed graph with road classes. The graph is written as follows:

$$ G=(R,E,L) $$

$$ R=\{r_i \mid i=1,2,\ldots,n\} $$

$$ E=\{r_{ij} \mid r_i\in R,\ r_j\in R,\ i\neq j\} $$

$$ L=\{r_{ij}^{l} \mid l=1,2,\ r_{ij}\in E\} $$

In this notation, each electric vehicle trip is a path in the graph. I do not assume that the shortest distance path is always the fastest path. Instead, I assign each edge a time-dependent impedance and solve the all-pairs shortest path problem with the Floyd algorithm. The Floyd recurrence I use is:

$$ d_{ij}^{(k)}=\min\left(d_{ij}^{(k-1)},\ d_{ik}^{(k-1)}+d_{kj}^{(k-1)}\right) $$

where d_ij^(k) is the shortest generalized travel time from node i to node j using at most the first k intermediate nodes. The generalized travel time includes edge travel time and node impedance. I reconstruct the optimal route from the predecessor matrix produced by the same recurrence. Because the electric vehicle load problem is repeated for many sampled electric vehicles, I cache the all-pairs shortest paths for each 30-minute traffic state and update only the edges whose flow-speed relation changes materially.

I represent the road topology in matrix form. For a small illustrative network, the adjacency matrix is:

$$ A=\begin{bmatrix} 0 & r_{12} & \infty & r_{14} & \infty & \infty & \infty & \infty & \infty \\ r_{21} & 0 & r_{23} & \infty & r_{25} & \infty & \infty & \infty & \infty \\ \infty & r_{32} & 0 & \infty & \infty & r_{36} & \infty & \infty & \infty \\ r_{41} & \infty & \infty & 0 & r_{45} & \infty & r_{47} & \infty & \infty \\ \infty & r_{52} & \infty & r_{54} & 0 & r_{56} & \infty & r_{58} & \infty \\ \infty & \infty & r_{63} & \infty & r_{65} & 0 & \infty & \infty & r_{69} \\ \infty & \infty & \infty & r_{74} & \infty & \infty & 0 & r_{78} & \infty \\ \infty & \infty & \infty & \infty & r_{85} & \infty & r_{87} & 0 & r_{89} \\ \infty & \infty & \infty & \infty & \infty & r_{96} & \infty & r_{98} & 0 \end{bmatrix} $$

Traffic flow is not fixed. I predict the 15-minute traffic flow on each road class with a wavelet neural network trained on historical flow, day type, and time-of-day features. The output is normalized and then converted into an average speed through a flow-speed relation. The relation I use is:

$$ \beta=a+b\left(\frac{q_{ij}(t)}{C_{ij}^{l}}\right)^{\eta} $$

$$ v_{ij}(t)=\frac{v_{l}^{0}}{1+\left(\frac{q_{ij}(t)}{C_{ij}^{l}}\right)^{\beta}} $$

Here v_l^0 is the free-flow speed of a class-l road, q_ij(t) is the real-time flow, and C_ij^l is the capacity. The coefficients a, b, and eta are calibrated by road class. As flow approaches capacity, speed falls nonlinearly. This matters for electric vehicle load prediction because a congested route increases travel time and energy consumption per unit distance.

The edge travel time follows directly from speed:

$$ T_{ij}(t)=\frac{r_{ij}}{v_{ij}(t)} $$

I also represent the delay at signalized intersections. The node impedance model I use is:

$$ T_C=0.441\frac{C_{ij}^{l}(1-\lambda)^2}{2(1-\lambda)\frac{q_{ij}(t)}{C_{ij}^{l}}+\left(\frac{q_{ij}(t)}{C_{ij}^{l}}\right)^2}+6.996 $$

where lambda is the green split at the intersection. The first term captures oversaturation delay, and the additive term represents fixed control delay. For a single trip with N road segments, the total travel impedance is:

$$ T_h=\sum_{n=1}^{N}T_{ij}^{n}+\sum_{n=1}^{N-1}T_C^{n,n+1} $$

The unit-distance energy consumption of an electric vehicle depends on speed and road class. I use a piecewise model. For class-I roads, the unit consumption is:

$$ d_e=0.247+\frac{1.52}{v_{ij}}-0.004v_{ij}+2.992\times10^{-5}v_{ij} $$

For class-II roads, the unit consumption is:

$$ d_e=-0.179+0.004v_{ij}+\frac{5.492}{v_{ij}} $$

The energy consumed on a segment is the product of unit consumption and distance:

$$ \Delta E_{ij}=d_e r_{ij} $$

I then update the state of charge after each segment:

$$ S_{t+1}=S_t-\frac{\Delta E_{ij}}{C_{bat}} $$

where C_bat is the battery capacity in kWh. This update is the bridge between the traffic model and the charging model. If the electric vehicle reaches a destination with sufficient state of charge, it parks and waits. If the state of charge falls below a minimum threshold, or if the remaining energy is insufficient for the next trip, the electric vehicle becomes a candidate for charging or swapping.

I now describe the multi-layer complex travel chain. I divide the urban area into four functional types: residential, workplace, commercial, and green space. A simple travel chain contains a single outbound and return trip between two functional areas. A complex travel chain contains multiple destinations and multiple purposes. The distinction is expressed as:

$$ \text{simple chain if } D_{l,n}=1 $$

$$ \text{complex chain if } D_{l,n}\gt 1 $$

where D_l,n is the number of destination layers in travel chain l at trip n. I generate the chain with Monte Carlo sampling. The initial location is sampled from the residential distribution during the morning peak and from the workplace or commercial distribution during the evening peak. The initial travel time is sampled from a mixture distribution that represents commuting, shopping, leisure, and shift work. The initial state of charge is sampled from a truncated normal distribution. The number of trips per day is sampled from a discrete distribution. The destination sequence is sampled from a purpose-dependent transition matrix. The dwell time at each destination is sampled from a lognormal distribution whose mean depends on the functional area and purpose.

Travel-chain and functional-area assumptions used for electric vehicle sampling
Functional area Main purpose Typical dwell-time pattern Charging preference Swap likelihood
Residential Rest, overnight parking, shift recovery Long evening and overnight dwell Slow charging is common; fast charging when dwell is short Low to medium
Workplace Work, meetings, daytime parking Long daytime dwell Slow charging during work hours; fast charging near departure Medium
Commercial Shopping, dining, leisure Short to medium dwell Fast charging dominates when dwell is short High when urgency is high
Green space Recreation, short stop Short dwell Opportunistic charging only Low

I model the travel chain as a sequence of states. The initial state is the origin and departure time. The first destination is sampled from the purpose-specific transition matrix. The arrival time is the departure time plus travel impedance. The dwell time is sampled. The next destination is sampled conditionally on the current functional area and the remaining planned activities. The chain terminates when the electric vehicle returns to the residential area or when the daily trip count is exhausted. I keep the chain spatially explicit by mapping each functional area to a set of road nodes and stations.

The charging model is based on state of charge. Let S_t^c be the state of charge of an electric vehicle connected to a charger at time t, C_bat be battery capacity, eta_c be charging efficiency, and P_t be charging power. The discrete-time update is:

$$ S_{t+1}^{c}=\frac{S_{t}^{c}C_{bat}+\eta_c P_t\Delta t}{C_{bat}} $$

I impose the following bounds and time-window constraints:

$$ S_{min}\le S_{t}^{c}\le S_{max} $$

$$ T_{begin}^{c}\le t\le T_{end}^{c} $$

$$ S_{end}=\eta_c S_{max} $$

Here S_min and S_max are the minimum and maximum allowable state of charge, T_begin^c and T_end^c are the connection and disconnection times, and S_end is the target state of charge at departure. In the simulation, a connected electric vehicle can choose slow charging or fast charging. Slow charging has low power and long duration. Fast charging has high power and short duration. I select the mode by minimizing a generalized cost that includes waiting time, charging time, and the risk of insufficient energy for the next trip. The cost function is:

$$ J_c=\alpha_w W_c+\alpha_t T_c+\alpha_r R_c $$

where W_c is waiting time, T_c is charging time, R_c is the remaining-range risk, and alpha_w, alpha_t, and alpha_r are weights. The remaining-range risk is:

$$ R_c=\max\left(0,\ E_{next}-(S_t-S_{min})C_{bat}\right) $$

where E_next is the energy required for the next trip. This cost function allows the model to choose fast charging when the dwell time is short and the next trip is long, and slow charging when the dwell time is long and the next trip is modest.

The swapping model is different. A battery swap replaces the depleted battery with a fully charged or nearly fully charged battery in a few minutes. I use the following target and constraints:

$$ S^{r}=0.9S_{max} $$

$$ S_{min}\le S_{t}^{r}\le S_{max} $$

$$ 0\lt t_{swap}\le 3\ \text{min} $$

After the swap, the electric vehicle leaves with S^r, while the depleted battery enters the station charging queue. The station operator must decide how to charge the depleted battery. A simple policy is to charge batteries in order of state of charge. That policy tends to produce more ready batteries quickly, but it can create an uneven state-of-charge distribution inside the station. I instead use a threshold-based queue. Batteries below a low threshold are fast-charged; batteries above the low threshold but below the target are slow-charged. This policy smooths the internal state-of-charge distribution and reduces abrupt station load spikes.

Battery-swapping queue policy used in the charge-swap interaction mode
Battery state of charge Queue class Charging mode Purpose
Below 0.20 Urgent Fast charging Raise low-SOC batteries quickly and reduce range anxiety
0.20 to 0.90 Normal Slow charging Fill the queue smoothly and shift load to lower-net-load periods
Above 0.90 Ready No charging Maintain a ready pool without unnecessary load

I write the battery charging power as a piecewise function of state of charge:

$$ P_b(S)=P_f \quad \text{if } 0\lt S\lt 0.2 $$

$$ P_b(S)=P_s \quad \text{if } 0.2\le S\lt 0.9 $$

The energy delivered to a depleted battery is the integral of charging power over the charging interval:

$$ E_b=\int_{t_b}^{t_b+\Delta t_b}P_b(S(t))dt $$

The station load at time t is the sum of the charging power of all batteries in the queue plus the power drawn by connected electric vehicles. This sum is what I aggregate in the spatiotemporal load map. The delay between the swap event and the battery charging event is the key mechanism that allows the swap station to shift load. In the charge-only mode, the electric vehicle begins charging immediately upon connection. In the charge-swap mode, the electric vehicle can leave quickly and the depleted battery can be charged later, subject to the queue policy and the available battery inventory.

The charge-swap interaction mode is a state-transition process. In the charging mode, the electric vehicle passes through driving, parking, connection, fast or slow charging, and departure. In the swapping mode, the electric vehicle passes through driving, arrival, battery exchange, and departure, while the station passes through depleted-battery admission, queueing, charging, and readiness. I represent the electric vehicle decision by a discrete choice among continue, slow charge, fast charge, and swap. The choice depends on the state of charge, remaining dwell time, next-trip energy, station queue length, and battery availability. A compact representation of the decision rule is:

$$ a^{*}=\arg\min_{a\in A}\left(J_{continue},J_{slow},J_{fast},J_{swap}\right) $$

where A is the action set. The continue action is feasible only if the state of charge is above the minimum and the remaining energy is sufficient for the next trip. The slow and fast charging actions are feasible only if a charger is available. The swap action is feasible only if a compatible battery is available and the station is open. The cost of swapping includes the swap fee, the travel time to the station, and the waiting time for a battery. The cost of charging includes the charging fee, the connection time, and the opportunity cost of parking.

For renewable generation, I model photovoltaic output with the single-diode engineering model. The current-voltage relation is:

$$ I=I_{SC}-C_1I_{SC}\left[\exp\left(\frac{U}{C_2U_{OC}}\right)-1\right] $$

$$ C_1=\left(1-\frac{I_m}{I_{SC}}\right)\exp\left(-\frac{U_m}{C_2U_{OC}}\right) $$

$$ C_2=\left(\frac{U_m}{U_{OC}}-1\right)\left[\ln\left(1-\frac{I_m}{I_{SC}}\right)\right]^{-1} $$

Here I is output current, U is output voltage, I_SC is short-circuit current, U_OC is open-circuit voltage, I_m is maximum-power-point current, and U_m is maximum-power-point voltage. The photovoltaic output is uncertain because irradiance and temperature vary. I represent the uncertainty by an interval:

$$ u_{PV}(t)\in\left[U(t)-\Delta u_{PV}^{max}(t),\ U(t)+\Delta u_{PV}^{max}(t)\right] $$

where u_PV(t) is the uncertain photovoltaic output, U(t) is the expected output, and Delta u_PV^max is the maximum allowable deviation. I use this interval in the comparison stage rather than in the deterministic load aggregation, so that the electric vehicle load can be evaluated against a range of renewable conditions.

For wind generation, I use the Weibull distribution for wind speed:

$$ f(v_w)=\frac{k}{c}\left(\frac{v_w}{c}\right)^{k-1}\exp\left[-\left(\frac{v_w}{c}\right)^{k}\right] $$

The wind turbine output is piecewise:

$$ P_{WT}=0 \quad \text{if } v_w\le v_{ci}\ \text{or}\ v_w\gt v_{co} $$

$$ P_{WT}=P_r\frac{v_w-v_{ci}}{v_r-v_{ci}} \quad \text{if } v_{ci}\lt v_w\le v_r $$

$$ P_{WT}=P_r \quad \text{if } v_r\lt v_w\le v_{co} $$

where v_ci is cut-in speed, v_r is rated speed, v_co is cut-out speed, and P_r is rated power. The wind and photovoltaic profiles are complementary: photovoltaic output is highest around midday, while wind output may be higher in the evening or early morning. The electric vehicle load, however, has its own peaks. The value of the charge-swap mode is that it can move part of the battery-charging load to periods when renewable output is high or net load is low.

Renewable and load interaction states used in the comparison
Period Photovoltaic availability Wind availability Electric vehicle load behavior Charge-swap effect
Early morning Low Moderate to high Low to moderate charging demand Depleted swap batteries can absorb low-net-load energy
Midday High Moderate Commercial fast charging and workplace charging Swap-station charging can be aligned with high photovoltaic output
Evening peak Low to zero Moderate Residential charging peak Swapping reduces immediate charging demand and lowers the peak
Late evening Zero Moderate Residential slow charging Deferred battery charging can fill the valley

I now describe the solution procedure. The procedure is Monte Carlo at the electric vehicle level and deterministic at the aggregation level. For each sampled electric vehicle, I generate a travel chain, sample an initial state of charge, simulate the route with the Floyd shortest path and the flow-speed relation, update the state of charge on each segment, and evaluate the charge-swap decision at each destination. I repeat this for the full electric vehicle fleet. I then aggregate the charging power and swapping-station power into 30-minute bins and map them to road nodes and functional areas. Finally, I subtract the renewable output from the load to obtain the net-load profile and compare the charge-only and charge-swap cases.

Solution procedure for the spatiotemporal charge-swap load prediction
Step Input Operation Output
1 Fleet size, electric vehicle penetration, battery capacity Sample initial location, initial time, initial SOC, and travel-chain length Electric vehicle agent set
2 Road graph, road class, capacity, free-flow speed Build adjacency matrix and run Floyd all-pairs shortest paths Time-dependent route set
3 Traffic flow forecast Apply flow-speed relation and node impedance Edge travel times and trip times
4 Speed, road class, distance Compute unit energy consumption and SOC decrement State of charge at each destination
5 SOC, dwell time, next-trip energy, station state Evaluate continue, slow charge, fast charge, and swap costs Charging or swapping action
6 Station queue and battery inventory Charge depleted batteries with threshold-based queue policy Swap-station load
7 Nodal charging and swapping power Aggregate into 30-minute bins by node and functional area Spatiotemporal load map
8 Photovoltaic and wind output Compute net load and compare charge-only and charge-swap modes Renewable accommodation indicators

I use a regional urban road network with 64 nodes and 98 road edges. Of these edges, 43 are class-I roads and 21 are class-II roads, with the remainder representing local connectors. The class-I free-flow speed is 60 km/h and the class-II free-flow speed is 40 km/h. I set the class-I capacity to 1560 PCU and the class-II capacity to 1450 PCU. The simulation horizon is 24 hours with a 30-minute aggregation step. The planned vehicle population is 70,000, and the electric vehicle penetration is 10 percent. This yields 7,000 electric vehicle agents. I distribute charging stations and swapping stations across the network nodes so that each functional area has access to both service types. The charging powers are 3.7 kW for slow charging and 30 kW for fast charging. The swapping station uses the same slow and fast charging powers for depleted batteries. A swap event itself takes no more than 3 minutes.

Simulation parameters for the electric vehicle charge-swap case study
Parameter Value Unit
Simulation horizon 24 h
Aggregation step 30 min
Road nodes 64 nodes
Road edges 98 edges
Class-I road edges 43 edges
Class-II road edges 21 edges
Planned vehicles 70,000 vehicles
Electric vehicle penetration 10 percent
Slow charging power 3.7 kW
Fast charging power 30 kW
Swap service time 3 min
Target swap state of charge 0.90 p.u.
Road-class parameters used in the flow-speed and energy-consumption models
Road class Free-flow speed Capacity Coefficient a Coefficient b Exponent eta
Class I 60 km/h 1560 PCU 1.726 3.15 3
Class II 40 km/h 1450 PCU 2.076 2.87 3

The initial travel time distribution is bimodal. The morning mode is centered on the commuting period, and the evening mode is centered on the return period. There is also a smaller midday mode for shopping and flexible work. The initial location distribution is weighted toward residential areas in the morning and toward workplaces and commercial areas in the afternoon. I do not force all electric vehicles to start from the same node. The spatial dispersion of origins is important because it changes the route length, the energy consumption, and the station choice.

I first evaluate the charging-only mode. In this mode, every electric vehicle that needs energy connects to a charger and remains connected until the target state of charge is reached or the dwell time expires. The aggregate charging load has three visible peaks. The first peak appears around 00:30, when late-returning electric vehicles connect for overnight slow charging. The second peak appears around 13:00, when commercial and workplace electric vehicles take advantage of midday parking. The third peak appears around 19:00, when residential charging coincides with the evening return period. The spatial distribution is uneven. Nodes serving residential areas carry the largest daily energy, followed by workplace nodes and commercial nodes. The uneven distribution is not an artifact. It follows from the functional mix, the number of service nodes, the dwell-time distribution, and the charging-mode choice.

Daily charging energy by functional area in the charging-only mode
Functional area Daily charging energy Dominant charging mode Main peak period Spatial pattern
Residential 460.6 MWh Slow charging with some fast charging Evening and overnight High load at many residential nodes
Workplace 217.8 MWh Slow charging during work hours Midday to afternoon Concentrated at employment nodes
Commercial 190.4 MWh Fast charging with short dwell Late afternoon and evening Moderate load at commercial nodes

The residential area has the largest daily charging energy because it has the largest population base, the widest service area, and the longest dwell time. Many electric vehicles arrive in the evening and remain connected overnight. The workplace area has the second-largest daily energy because workday parking creates long, predictable charging windows. The commercial area has the smallest daily energy among the three main functional types, but it has a higher fast-charging share because dwell times are short. This is an important distinction: daily energy and peak power are not the same. The commercial area can have a relatively small daily energy but a sharp local peak because many electric vehicles request fast charging within a short window.

I compare the charging-only mode with the charge-swap interaction mode. In the charge-swap mode, an electric vehicle can swap its battery when the expected charging time is incompatible with the remaining dwell time or when the next trip requires more energy than the charger can deliver before departure. The depleted battery then enters the station queue. The station uses the threshold-based policy described earlier. The total load is the sum of connected electric vehicle charging and station battery charging. The resulting profile is smoother than the charging-only profile. The peak reduction is not uniform across space. Residential nodes benefit most because overnight charging can be deferred and because the station can spread battery charging across the night. Commercial nodes benefit in the short-dwell periods because swapping removes urgent fast-charging demand from the immediate peak.

Qualitative comparison of charging-only and charge-swap modes
Indicator Charging-only mode Charge-swap interaction mode Interpretation
Peak charging power Higher Lower Swapping removes part of the immediate charging demand
Valley filling Limited Stronger Depleted batteries can charge during low-net-load periods
Load variance Higher Lower Deferred battery charging smooths the profile
Renewable accommodation Moderate Higher Battery charging can follow photovoltaic output
Range anxiety Higher for short-dwell users Lower Swap provides a fast energy supplement
Station complexity Lower Higher Battery inventory and queue management are required

The spatial behavior of the swap load is also informative. Swap-station load is generally lower than charging-station load because the swap station serves as a supplement rather than the primary energy source. The swap load is concentrated at nodes that connect residential, workplace, and commercial trips. In time, the swap load is more evenly distributed than the charging load. There are mild concentrations in the late morning and mid-afternoon, but the sharp evening peak of the charging-only mode is weakened. This is exactly the compensation mechanism I want to capture: battery swapping does not eliminate electric vehicle load, but it changes the timing of that load and reduces the coupling between trip urgency and instantaneous charging demand.

Representative spatial distribution of swap-station load
Spatial cluster Functional context Swap-load pattern Reason
Residential-adjacent nodes Residential and mixed use Moderate overnight and early-morning load Depleted batteries from evening returns are charged after the peak
Employment-adjacent nodes Workplace and transit Moderate midday load Workday electric vehicles use swapping when charging bays are occupied
Commercial-adjacent nodes Commercial and leisure Short, high-urgency swaps Short dwell times make swapping attractive
Connector nodes Mixed corridor Low but steady load Vehicles use swapping opportunistically along the route

I next compare the load profiles against renewable generation. Photovoltaic output rises from morning to midday and falls in the afternoon. Wind output varies more smoothly but is not guaranteed to coincide with the electric vehicle peak. In the charging-only mode, the evening residential peak often occurs when photovoltaic output is low or zero. This increases the net-load peak and may require additional flexible resources. In the charge-swap mode, part of the depleted-battery charging can be moved to midday or early afternoon, when photovoltaic output is high. The swap station therefore acts as a small, distributed energy buffer. It does not generate energy, but it shifts the time at which the distribution network supplies that energy.

I quantify the effect with three indicators. The first is peak net load:

$$ L_{peak}^{net}=\max_{t}\left(L_{EV}(t)-P_{PV}(t)-P_{WT}(t)\right) $$

The second is valley net load:

$$ L_{valley}^{net}=\min_{t}\left(L_{EV}(t)-P_{PV}(t)-P_{WT}(t)\right) $$

The third is the net-load variance:

$$ \sigma_{net}^{2}=\frac{1}{T}\sum_{t=1}^{T}\left(L_{EV}(t)-P_{PV}(t)-P_{WT}(t)-\bar{L}_{net}\right)^{2} $$

A lower peak net load and a lower variance indicate better smoothing. A higher valley net load can indicate better valley filling, provided that the valley is not pushed into reverse flow. In the case study, the charge-swap mode reduces the evening peak more than it increases the midday load. The reason is that the swap station does not charge every depleted battery immediately. The threshold queue delays part of the charging until the station has spare capacity or until the net load is lower. This deferral is limited by battery inventory and by the number of ready batteries, so the swap station cannot shift load indefinitely. The queue policy must therefore balance service reliability and load smoothing.

Renewable interaction comparison between the two modes
Metric Charging-only mode Charge-swap mode Expected direction
Evening net-load peak High Lower Improved
Midday net-load valley Deep Less deep Improved if reverse flow is avoided
Net-load variance High Lower Improved
Photovoltaic self-consumption Moderate Higher Improved
Battery inventory pressure Not applicable Moderate to high Requires management
Service waiting time Charging queue dependent Swap queue dependent Requires station sizing

The functional-area comparison shows that the electric vehicle charging load is not a single curve. Residential load is dominated by long dwell and slow charging. Workplace load is dominated by daytime parking and predictable charging windows. Commercial load is dominated by short dwell and fast charging. These differences explain why a uniform charging tariff or a uniform station-sizing rule would be inefficient. A residential station needs many slow chargers and a modest number of fast chargers. A commercial station needs a higher fast-charging share and a larger incoming transformer capacity relative to its daily energy. A workplace station can use slow charging during work hours and can coordinate with photovoltaic output if parking is exposed to the sun. A swapping station can be placed at the boundary between functional areas, where it can serve multiple trip purposes and absorb deferred battery charging.

Functional-area design implications for electric vehicle energy supply
Functional area Charging capacity need Swapping role Renewable coordination Operational priority
Residential Many slow chargers, some fast chargers Secondary; useful for late returns and urgent morning trips Overnight wind and valley filling Avoid creating a new evening peak
Workplace Moderate slow charging, limited fast charging Moderate; useful when charging bays are full Midday photovoltaic consumption Coordinate workday charging with solar output
Commercial High fast-charging share High; short dwell makes swapping attractive Limited direct solar matching unless parking is shaded or covered Reduce queuing and local peak power
Green space Low Low Opportunistic Maintain low utilization and avoid oversizing

I also examine the sensitivity of the results to the swap threshold. If the threshold is too low, most depleted batteries enter fast charging, and the swap station behaves like a fast-charging station with a delay. If the threshold is too high, too many batteries enter slow charging, and the station may not maintain enough ready batteries during high-demand periods. The best threshold depends on the arrival rate, the battery inventory, the fast-charging power, the slow-charging power, and the service-level requirement. I represent the trade-off with a simple service reliability index:

$$ R_{service}=\frac{N_{ready}}{N_{demand}} $$

where N_ready is the number of batteries ready at the time of demand and N_demand is the number of swap requests. A higher R_service improves user satisfaction but may increase the station’s immediate charging load. A lower R_service reduces immediate load but increases waiting time and the risk of failed swaps. The threshold policy is therefore a compromise between electric vehicle service quality and grid-side load smoothing.

Trade-off between swap-queue threshold and system performance
Threshold setting Fast-charging share Ready-battery reliability Peak load effect Operational risk
Very low High High in the short term Weak peak reduction High station peak power
Moderate Balanced Acceptable Moderate to strong peak reduction Manageable with inventory control
Very high Low Potentially low Strong load shifting Battery shortage and long waits

The model also reveals a spatial equity issue. Stations located in high-demand commercial nodes may face many short-dwell swap requests, while stations located in residential nodes may face many overnight charging requests. If the same battery inventory and queue policy are applied everywhere, some stations will be overutilized and others underutilized. A more efficient design would allow battery sharing among nearby stations or dynamic repositioning of charged batteries. I do not optimize this repositioning in the present model, but the results show why it matters. The swap station is not only an electric vehicle service point. It is also a distributed storage-like resource whose value depends on its location and its ability to defer charging.

I further analyze the effect of electric vehicle penetration. At low penetration, the charging load is small relative to the base load, and the difference between charging-only and charge-swap modes is modest. At higher penetration, the evening peak becomes more pronounced, and the swap mode becomes more valuable. The relationship is nonlinear because station utilization, queue length, and battery inventory interact. I express the penetration sensitivity as:

$$ L_{EV}(t)=\sum_{i=1}^{N_{EV}}P_i(t) $$

$$ N_{EV}=\rho_{EV}N_{vehicle} $$

where rho_EV is the electric vehicle penetration and N_vehicle is the total vehicle population. As rho_EV increases, the variance of the aggregate load increases faster than the mean if charging is uncontrolled. The swap mode reduces this variance by decoupling the electric vehicle departure time from the battery charging time. The decoupling is limited by the number of spare batteries, but even a modest spare inventory can significantly reduce the evening peak.

Expected penetration sensitivity of the charge-swap benefit
Electric vehicle penetration Charging-only peak Charge-swap peak Relative benefit Main constraint
Low Small Small Limited Station utilization is low
Medium Moderate Lower Visible Battery inventory and queue management
High High Materially lower Strong Station capacity and dynamic battery repositioning

I also consider the uncertainty of renewable output. Photovoltaic output can drop sharply because of clouds, and wind output can deviate from the forecast. The charge-swap mode is robust to this uncertainty only if the station has enough battery inventory and enough flexible charging capacity. If the station commits too much load to midday photovoltaic output and the output falls, the station may need to import power during the evening peak. I therefore treat the swap station as a flexible load rather than a guaranteed renewable sink. The correct operational rule is to charge depleted batteries when the net load is low and to reduce charging when the net load is high, subject to the ready-battery constraint. This rule can be written as:

$$ P_{b}^{charge}(t)=\min\left(P_{b}^{max},\ \max\left(0,\ P_{b}^{ref}(t)-\gamma L_{net}(t)\right)\right) $$

where P_b^ref is the reference charging power, L_net is the net load, and gamma is a smoothing coefficient. The min and max operators enforce power limits and prevent negative charging. This rule is a simple form of net-load-aware battery charging. It can be extended with price signals, carbon intensity signals, or transformer loading constraints.

The model has several limitations that I state clearly. First, I simulate station availability but do not optimize station siting. The nodal load differences therefore reflect the assumed station distribution as well as the travel demand. Second, I use a threshold-based battery queue rather than a full stochastic inventory optimization. A more advanced model could optimize the number of batteries, the charging schedule, and the service level jointly. Third, I model traffic flow with a wavelet neural network and a flow-speed relation, but I do not model individual lane changes, signal coordination, or incidents. Fourth, I represent renewable uncertainty with an interval rather than a full scenario tree. A robust or stochastic optimization could be added to the dispatch layer. Fifth, I do not model battery degradation explicitly. Fast charging and deep discharging can affect battery health, and a full life-cycle assessment would need to include degradation costs.

Despite these limitations, the framework provides a coherent way to connect electric vehicle travel behavior, traffic conditions, charging decisions, battery swapping, and renewable generation. The key insight is that the electric vehicle load is not only a function of how many electric vehicles exist. It is a function of where they go, how long they stay, how much energy they need for the next trip, and how quickly they can obtain that energy. Battery swapping changes the timing of energy delivery. It does not remove the energy requirement, but it separates the service event from the charging event. This separation is valuable because it gives the distribution network more freedom to choose when to supply energy.

I summarize the main findings as follows. The electric vehicle charging load has clear regional differences. Residential areas dominate daily energy because of long overnight dwell and large population base. Workplace areas have a predictable daytime load. Commercial areas have smaller daily energy but sharper fast-charging peaks because dwell times are short. The charge-swap interaction mode reduces the immediate peak by allowing depleted batteries to be charged later. The swap-station load is lower than the charging-station load and is more evenly distributed in time. With renewable generation, the swap mode improves valley filling and increases the opportunity to use midday photovoltaic output. The benefit depends on the battery inventory, the queue threshold, and the station location. The model also shows that the swap station should not be treated as a passive load. It should be managed as a flexible, storage-like resource whose charging can be shifted within the limits of service reliability.

Summary of the main electric vehicle charge-swap findings
Finding Evidence in the model Implication for distribution networks
Regional load differences are significant Residential, workplace, and commercial areas have different daily energy and peak timing Station sizing and tariffs should be area-specific
Swapping reduces the immediate peak Depleted batteries are charged after the swap event Swap stations can provide peak-shaving flexibility
Swap load is more evenly distributed Threshold queue and battery inventory spread charging over time Transformer loading can be smoothed
Renewable accommodation improves Battery charging can be aligned with photovoltaic output Charge-swap coordination can support renewable integration
Service reliability constrains flexibility Too much deferral reduces ready-battery availability Battery inventory and queue policy must be co-designed

For future work, I plan to extend the model in three directions. First, I will integrate station siting and battery inventory optimization so that the spatial distribution of charging and swapping stations is determined endogenously. Second, I will add battery degradation and fast-charging stress so that the trade-off between service speed and battery life is represented explicitly. Third, I will formulate a two-stage robust dispatch model that uses the predicted electric vehicle load and renewable uncertainty to schedule battery charging, station inventory, and network flexibility. These extensions would move the framework from prediction toward operational decision support.

In conclusion, I have presented a multi-layer complex travel-chain model for electric vehicle charging and battery-swapping load prediction. The model uses the Floyd algorithm for route choice, a flow-speed relation for traffic-dependent energy consumption, a charge-swap interaction mode for station choice, and wind-photovoltaic models for renewable comparison. The results show that electric vehicle charging demand is spatially and temporally heterogeneous, and that battery swapping can effectively compensate for charging-load fluctuations when it is managed with a threshold-based queue and sufficient battery inventory. The framework provides a practical basis for evaluating how electric vehicle energy supply, distribution network operation, and renewable accommodation interact under realistic urban travel behavior.

Scroll to Top