Virtual Battery Digital Twin Evaluates Electric Vehicle Charging Demand and Storage

Analyzing 348,404 screened electric vehicle charging sessions across Jiaxing, China, researchers have established that combining machine learning load forecasts with an offline decision twin can reduce modeled daily charging station electricity costs by 3.31 percent and peak demand by 3.14 percent under demand uncertainty. The empirical investigation, conducted by Zhengyang Fang of Ningbo University of Finance and Economics and published in Frontiers in Mechanical Engineering, evaluated whether transaction records alone could support pre-deployment battery energy storage scheduling when real-time charger telemetry and physical on-site batteries are entirely absent. By converting 6.754 gigawatt-hours of historical charging energy from 13 public stations into 30-minute demand profiles, the investigation tested how gradient-boosted decision trees, empirical quantile models, and split-conformal calibration perform when paired with linear programming dispatch. Rather than assuming idealized operational control or asserting universal forecasting dominance, the study demonstrated that nonlinear predictive algorithms yield a modest but statistically stable 1.65 percent root mean squared error reduction over a simple seven-day seasonal persistence benchmark, while providing an auditable pre-deployment framework for sizing and operating virtual stationary batteries.

The study began by processing a public municipal dataset consisting of 441,077 raw charging records collected from 92 anonymized charging posts across three administrative districts in Jiaxing between December 31, 2019, and December 31, 2021. To ensure rigorous data integrity, prespecified screening criteria eliminated corrupted logs, retaining records with verified zero abnormality flags, strictly positive delivered energy, valid connection timestamps, and total durations spanning between one minute and twelve hours. This screening retained 348,404 sessions, representing 79.0 percent of raw transaction records and comprising 52,581 anonymous charging users. Across these retained sessions, the median delivered energy was 17.64 kilowatt-hours and the median connection duration was 40.23 minutes. The median session-average power measured 26.07 kilowatts, with a 95th percentile of 46.98 kilowatts. Because the municipal dataset omitted specific hardware classifications regarding whether individual posts utilized alternating-current level-two or direct-current fast-charging technology, session-average power values were treated strictly as complete session summaries rather than operational charger ratings, preventing improper technological inferences from skewing analytical parameters.

Because raw transaction logs record only aggregate session energy rather than continuous real-time telemetry, the researcher reconstructed interval demand by allocating each session’s energy uniformly across its active 30-minute intervals. Summing these overlapping sessions across half-hour increments yielded continuous station demand while conserving the screened volume of 6,753,960.32 kilowatt-hours. The resulting time series revealed substantial intraday patterns, with station demand falling to minimums overnight, climbing throughout the morning hours, and peaking near 51.4 kilowatts around 12:30 on weekdays, with weekend peaks occurring slightly later. To directly address potential reconstruction bias, the author constructed two alternative, energy-conserving profiles: a front-loaded profile declining linearly from 1.5 to 0.5 times average session power, and a back-loaded profile reversing that weighting. When refitting the primary forecasting model and re-evaluating deterministic storage dispatch at Bus Station, cost reductions shifted by at most 0.03 percentage points and peak reductions by at most 0.08 percentage points. This confirmed that while within-session load shape affects interval forecasts, modeled deterministic storage economic outcomes remain robust to stylized energy-conserving profile shifts.

The forecasting framework evaluated point and probabilistic day-ahead load predictions over an untouched holdout period from October 1 to December 31, 2021, comprising 52,974 half-hour observations across 12 operational stations. Technology Park was omitted from holdout testing because its records terminated on September 28, 2021, avoiding synthetic data imputation. Training on 394,145 historical rows through September 30, 2021, the point LightGBM model configured with 220 boosting trees, 31 leaves, and a 0.05 learning rate achieved a root mean squared error of 29.37 kilowatts, a mean absolute error of 21.20 kilowatts, a weighted absolute percentage error of 45.49 percent, and a coefficient of determination of 0.647. An independent XGBoost baseline reached an RMSE of 29.52 kilowatts, whereas a simple seven-day same-slot mean baseline achieved 29.86 kilowatts, and a single-day seasonal naive model lagged at 38.33 kilowatts. Resampling 1,104 station-day clusters across 10,000 bootstrap iterations revealed that LightGBM’s 0.493 kilowatt advantage over the seven-day baseline possessed a 95 percent confidence interval of 0.287 to 0.701 kilowatts, confirming a reliable 1.65 percent relative gain.

Station-by-station evaluations exposed significant performance divergence across geographic locations that pooled metrics obscured. LightGBM produced superior root mean squared error metrics at 10 of the 12 evaluation stations, generating its highest relative improvement of 9.44 percent at Park A, where RMSE dropped from 21.26 kilowatts under the seven-day baseline to 19.25 kilowatts. At Bus Station, the relative gain was 1.02 percent, with RMSE decreasing from 36.38 to 36.01 kilowatts. However, the machine learning model degraded performance at two facilities: root mean squared error increased by 2.61 percent at Park B, rising from 40.70 to 41.76 kilowatts, and worsened by 6.74 percent at Financial Industrial Park, where error climbed from 22.52 to 24.03 kilowatts. Because all 12 test stations were present in the training corpus, the author noted that these findings demonstrate temporal generalization across known heterogeneous stations rather than successful cold-start transfer to entirely novel, unobserved charging sites.

Feature attribution via TreeSHAP on 6,000 sampled holdout observations clarified why complex gradient boosting yielded limited improvements over basic seasonal persistence. The seven-day same-slot mean demand emerged as the overwhelming predictor, commanding a mean absolute SHAP value of 22.60 kilowatts. Previous-day mean station demand ranked second at 4.66 kilowatts, followed by cyclic diurnal harmonic terms at 3.93 and 2.38 kilowatts. Exogenous weather metrics—including temperature, relative humidity, and precipitation aggregated from 2,196 district-date weather records—alongside time-of-use electricity pricing tiers contributed negligible predictive power. This attribution showed that LightGBM primarily refined localized daily and weekly habit persistence rather than uncovering complex meteorological drivers. Station demand uncertainty was modeled through dedicated LightGBM quantile regression optimizing pinball loss for the 5th and 95th percentiles. The uncalibrated quantile envelope averaged 83.16 kilowatts in width but covered only 79.56 percent of test observations. When an independent split-conformal calibration was executed on September 2021 residuals, pooled holdout coverage rose to 89.03 percent with an average interval width of 76.86 kilowatts, though individual station coverage continued to range between 83.79 and 91.89 percent.

To evaluate storage dispatch, an offline decision twin simulated a virtual battery energy storage system at Bus Station, which provided 53,346 retained sessions and an unbroken 92-day test sequence. The modeled storage unit was scaled to the station’s 95th percentile training demand of 158.31 kilowatts, yielding a rated power of 79.15 kilowatts and an energy capacity of 158.31 kilowatt-hours for a two-hour duration. The simulation enforced 95 percent one-way charge and discharge efficiencies, stored energy boundaries restricted between 10 and 90 percent of total capacity, initial and terminal states fixed at 50 percent, and an absolute prohibition against reverse grid export. Multi-objective trade-offs between daily electricity procurement expenditure under local time-of-use tariffs—ranging from 0.3784 to 1.2064 Chinese yuan per kilowatt-hour—and maximum half-hour grid power draw were mapped using the epsilon-constraint linear programming method. Compromise dispatch schedules were identified by calculating the minimum Euclidean distance to the normalized cost-peak utopia point, treating grid capacity and energy savings as explicit competing priorities without arbitrarily monetizing capital depreciation.

Over the 92-day holdout sequence, deterministic scheduling based on the central point forecast reduced modeled daily electricity expenses from 1,555.85 to 1,503.59 yuan, representing a 3.52 percent financial saving, while trimming peak grid demand by 7.27 percent from 130.95 to 122.00 kilowatts, operating at an intensity of 0.632 equivalent full cycles per day. When optimizing across the three-trajectory uncertainty envelope encompassing the 5th percentile, central path, and 95th percentile demand forecasts, the system reduced expected daily power costs by 3.31 percent, lowering expenditures from 1,519.64 to 1,469.44 yuan. The robust three-trajectory dispatch simultaneously lowered 95th-percentile peak grid draw by 3.14 percent, shifting maximum grid intake from 189.16 to 183.22 kilowatts with an average cycling rate of 0.560 equivalent full cycles daily. The smaller peak shaving observed in the robust model reflected the mathematical requirement that a single, nonanticipative battery schedule remain strictly feasible across extreme high and low demand realizations simultaneously, representing an explicit operational insurance cost.

The study author emphasized that all reported operational benefits remain conditional simulation outcomes from an offline decision-support model rather than realized empirical savings from physical hardware. The offline decision twin tracks system state transitions, battery dynamics, and energy balance equations, but lacks live telemetry synchronization, automated state-of-health estimation, communication latency management, and receding-horizon closed-loop control. On a representative test day reaching an observed peak demand of 239.44 kilowatts, realized load exceeded the 95th percentile trajectory during one interval, illustrating that statistical envelopes cannot provide absolute coverage guarantees against unexpected demand spikes. Furthermore, sensitivity evaluations varying storage power from 0.25 to 0.75 times peak demand and duration from one to four hours demonstrated non-linear diminishing returns without establishing commercial sizing recommendations, as battery capital costs and electrochemical degradation were excluded from available records. The findings establish that municipal transaction logs support disciplined pre-deployment feasibility screening, but physical deployment requires site-specific interconnection agreements, live charger telemetry, and verified battery management systems.

Subject of Research: Pre-deployment energy storage scheduling and machine-learning load forecasting for electric vehicle charging stations using transaction records

Article Title: Data-driven digital twin-assisted energy storage scheduling for electric vehicle charging stations under demand uncertainty: explainable forecasting and multi-objective dispatch

Article References: Fang, Z. (2026). Data-driven digital twin-assisted energy storage scheduling for electric vehicle charging stations under demand uncertainty: explainable forecasting and multi-objective dispatch. Frontiers in Mechanical Engineering, 12, Article 1986327. https://doi.org/10.3389/fmech.2026.1986327

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1986327

Keywords: electric vehicles, energy storage, digital twin, LightGBM, load forecasting, battery energy storage systems, demand uncertainty, charging infrastructure, Data-driven, digital, twin-assisted, energy