Dynamic Machine Learning-Based Simulation for Preemptive Supply-Demand Balancing Amid EV Charging Growth in the Jamali Grid 2025–2060

Abstract

The rapid uptake of electric vehicles (EVs) in the Jawa–Madura–Bali (Jamali) grid produces highly variable charging demands that threaten the supply–demand balance. To forestall instability, we developed a predictive simulation based on long short-term memory (LSTM) networks that combines historical generation and consumption patterns with models of EV population growth and initial charging-time (ICT). We introduce a novel supply–demand balance score to quantify weekly and annual deviations between projected supply and demand curves, then use this metric to guide the machine-learning model in optimizing annual growth rate (AGR) and preventing supply demand imbalance. Relative to a business-as-usual baseline, our approach improves balance scores by 64% and projects up to a 59% reduction in charging load by 2060. These results demonstrate the promise of data-driven demand-management strategies for maintaining grid reliability during large-scale EV integration.

1. Introduction

The integration of electric vehicles (EVs) into Indonesia’s power grid presents both opportunities and challenges in achieving a sustainable and stable energy transition. While EV adoption contributes to reduced fossil fuel dependency and lower greenhouse gas (GHG) emissions, it also introduces new demand-side complexities to the electricity grid [1]. Unlike conventional loads, EV charging demand is highly intermittent, influenced by charging behavior, vehicle ownership trends, and urban traffic conditions. Without proactive supply-demand balancing, widespread EV adoption could lead to significant grid instability due to unsynchronized demand surges [2].

This study introduces a machine learning-based predictive simulation to model EV-driven demand growth within the Jawa-Madura-Bali (Jamali) power grid, Indonesia’s most heavily loaded electrical system [3]. The simulation forecasts supply-demand imbalances and recommends corrective actions to prevent demand overshoot while optimizing annual growth rate (AGR) and vehicle preference shifts. This simulation enables proactive grid planning, offering a data-driven approach for policymakers to align EV adoption with energy system constraints.

1.1. Jawa Madura Bali Power System and Its Supply Demand Historical Growth

Indonesia’s power system consists of several interconnected grids, which are vital for managing the diverse energy needs across the archipelago. The country’s grid infrastructure is divided into multiple regions, with major grids including Sumatra, Jawa-Bali, Kalimantan, Sulawesi, and Papua. Among these, the Jamali grid stands out as the largest and most significant, both in terms of capacity and demand [4].

The Jamali grid serves the densely populated Java Island, which is home to approximately 55% of Indonesia’s population with most vehicles and hosts the capital city, Jakarta [5]. This high population density, coupled with economic activity, places considerable demands on the grid. Java island, being the most inhabited region in Indonesia, is central to the country’s power system. The Jamali grid is an extensive interconnected system that manages both supply and demand across the islands of Java, Madura, and Bali. Historically, the Jamali grid has experienced significant growth in both supply and demand [6]. The demand has surged due to increasing urbanization, industrialization, and a rise in residential consumption, while the supply side has been driven by substantial investments in power generation and infrastructure upgrades.

In recent years, the Jamali grid has faced evolving challenges due to the growing integration of intermittent renewable energy sources and the increased demand from EV charging [7]. The anticipated surge in EV usage is expected to further influence the grid’s supply-demand balance. Understanding these historical trends and current developments is crucial for ensuring a stable and reliable power system for the future [8].

The Jamali grid’s supply side is supported by a network of diverse power plants, including thermal, hydroelectric, and increasingly renewable energy sources. As of 2023, the Jamali grid comprises 402 power plants with a total installed capacity of around 48,000 megawatts (MW). This capacity is distributed among several large thermal power plants, including coal and gas-fired plants, as well as a growing number of hydro and solar installations [9]. The grid’s supply capacity has seen consistent growth over the years, driven by ongoing investments and upgrades aimed at meeting the rising energy demand and incorporating more sustainable energy sources.

Figure 1 illustrates the historical growth in supply capacity for Indonesia’s power system (1974–2004), as reflected in the Jamali grid.

The peak demand has risen sharply; introducing EV fleets without corrective measures could further strain the balance.

Table 1. Installed capacity and peak demand load of Jamali grid.

Year	Total Installed Capacity in MW	Peak Demand Load in MW
2014	31.062	23.908
2015	27.867	24.269
2016	36.712	33.208
2017	36.517	26.580
2018	37.721	27.097
2019	40.174	26.657
2020	40.685	24.420
2021	41.743	25.852
2022	45.835	24.228
2023	47.647	40.223
2024	50.103	42.635

presents the installed capacity and peak demand of the Jamali grid over the past ten years.

1.2. Development of Electric Vehicle Growth in Indonesia and the Impact on the Grid

The adoption of EVs in Indonesia is in its early stages but is set for significant expansion. As of 2023, the number of EVs in Indonesia ranges between 10,000 and 20,000, marking the beginning of a broader market development [10]. The Indonesian government aims to escalate this number to 2 million EVs by 2030 [11] through a variety of incentives. These include tax breaks, reduced import duties, and subsidies aimed at both EV purchases and the development of essential charging infrastructure.

Indonesia has implemented a range of regulations and incentives to promote the adoption of EVs as part of its strategy to address environmental and energy challenges. Key measures include the odd-even license plate policy in major cities like Jakarta, which restricts vehicle use based on license plate numbers on certain days to reduce traffic congestion and air pollution, thereby indirectly encouraging EV adoption [12]. Additionally, the government offers tax incentives, such as reduced import duties and exemptions from luxury goods taxes, making EVs more affordable and stimulating market growth [13]. Further incentives include free or discounted parking for EVs in some cities, which lowers the overall cost of ownership and enhances consumer appeal [14].

To support the expanding EV market, the government is also investing in charging infrastructure, with initiatives to install charging stations in public areas and along highways [15]. Additionally, the Indonesian government has introduced further subsidy programs to accelerate the transition to electric vehicles. These programs include a 10% VAT reduction for four-wheeled electric vehicles (4wEVs), as stipulated in Ministry of Finance Regulation No. 38 of 2023 [16], and a deduction of IDR 7,000,000 for two-wheeled electric vehicles (2wEVs), as detailed in Ministry of Industry Regulation No. 6 of 2023 [17]. The political landscape in Indonesia is a critical factor in shaping the growth of EVs within the country.

At present, the impact of EVs on the grid is relatively minor due to their limited numbers [18]. However, as EV adoption grows, particularly during peak charging periods, there will be increased demand on the grid.

1.3. Literature Review and Overview of Contributions

This study extends our earlier MDPI publication, Simulating EV Growth Scenarios in Jawa–Madura–Bali from 2024 to 2029: Balancing the Power Grid’s Supply and Demand, by integrating advanced machine learning methods to project grid dynamics through 2060.

Existing work on EV–grid integration has predominantly employed static supply–demand-coupled frameworks, exemplified by Song and Hu’s learning-based optimization of charging station locations for demand management [19] and Tang et al.’s dual-layer MPC for coordinated battery load demand and grid-side supply matching at swapping stations [20]. However, these static frameworks often underestimate peak deficits and overlook evolving temporal charging patterns, as demonstrated in comparative analyses of dynamic versus predetermined load scenarios.

To enhance forecast accuracy, recent studies have leveraged recurrent architectures, particularly long short-term memory (LSTM) networks, where Liu et al. developed a multiheaded convolutional LSTM model for green energy forecasting, improving solar irradiance and wind power predictions by over 25% [21], and Eze and Ayorinde’s systematic review confirmed that LSTM-based approaches consistently outperform traditional methods, reducing forecasting errors under variable conditions [22].

Additionally, the role of initial charging time (ICT) distributions in shaping aggregate EV demand has been quantified: Wang et al. used Monte Carlo simulation on charging-station order data to show that optimized ICT profiles can lower peak loads by 15–20% [23], while Wan et al. introduced a model-free deep reinforcement learning scheduler that adaptively learns real-time strategies, achieving up to 20% load flattening under dynamic pricing [24]. Collectively, these findings advocate for integrated frameworks that combine advanced LSTM forecasting with ICT-aware scheduling and reinforcement-learning control to ensure resilient and efficient EV–grid integration under rising penetration rates.

Table 2. Literature review of current existing LSTM and EV related studies.

Author(s)	Year	Focus Area	Grid Aspect Focus	EV Aspect Focus	Novelty/Limitation
Cao et al. [25]	2019	Microgrid	Addition of BESS	Even distribution of EVs	Unlimited supply from macrogrid
Fayyazbakhsh et al. [26]	2025	Central Europe	Use of dummy variables	EV load mixed with total load	EV load not specifically modeled
Afifi et al. [27]	2025	IEEE Bus Networks	Focus on power loss	Uses daily and monthly patterns	No modeling of EV population growth
Guldorum et al. [28]	2022	EV parking area	Focus on peak load limitation	Focus on V2G	Static EV load
Feng et al. [29]	2024	Driving cycle scenario	Not the main focus	Dynamic simulation of EV power	Static proportions of EV types
Li et al. [30]	2020	Residential area	Focus on improving efficiency and resilience	Scheduled V2G scenarios	Single EV charging preference
Tian et al. [31]	2021	Palo Alto	Uses seasonal patterns	Uses general load curves	Static and periodic EV load
Tampubolon et al. (this study)	2025	Java-Madura-Bali	Historical data simulation and AGR growth prediction	Simulation of EV population, preference, and load patterns	Considers dynamic load by each EV type

shows the positioning of this research study in the context of existing LSTM-based energy and EV grid prediction literature.

Building on these insights, the present study contributes to the field in five key ways:

• It introduces a machine learning-driven supply–demand balance simulation that dynamically adjusts EV charging profiles across multiple scenarios, replacing static assumptions with data-informed trajectories.
• It leverages LSTM networks to forecast long-term supply and demand growth, enabling the recommendation of corrective actions when projected demand growth threatens grid stability.
• It implements a suite of EV adoption scenarios—incorporating shifts in vehicle preferences, ICT optimization, and load-reduction strategies—to reflect plausible technology and behavioral trends.
• It develops a quantitative scoring system to evaluate grid stability under each scenario, thereby assessing the effectiveness of machine learning-informed interventions.
• It demonstrates that machine learning-based optimization of EV growth can significantly improve balance scores relative to business-as-usual projections, ensuring more reliable integration of electric vehicles into the Jawa–Madura–Bali grid.

Together, these contributions establish a robust, data-driven framework for evaluating policy and infrastructure strategies in support of Indonesia’s low-carbon transport transition.

1.4. Structure of the Paper

The remainder of this paper is structured as follows:

• Section 2 (Materials) details the machine learning methodology, data sources, and predictive modeling techniques used in the simulation.
• Section 3 (Methods) presents the supply-demand simulation framework, including demand growth factors, ICT modeling, and scenario configurations.
• Section 4 (Simulation Results) presents the datasets for the LSTM train, validation test, 2024 input, 2029 BAU output, and the output 2060 (BAU and if corrective measures taken).
• Section 5 (Discussions) provides discussions, comparing ML-predicted outcomes against BAU simulations and analyzing policy-driven impacts on grid stability.
• Section 6 (Conclusion) concludes with key insights, policy recommendations, and future research directions for EV integration in Indonesia’s power grid.

2. Materials

2.1. Understanding EV Intermittency

Electric Vehicle intermittency refers to the variability in energy demand from EVs due to factors like charging patterns and usage. Managing this intermittency is crucial for ensuring a stable energy supply and efficient grid operation. Following is a detailed overview that combines technical calculations with strategies for integrating EV demand.

As the number of EVs grows, total energy demand increases, especially during charging periods when energy consumption is at its peak. This rise in demand is particularly pronounced when many EVs charge simultaneously or during peak hours, leading to significant peaks in energy demand [7]. In the context of EVs, ICT refers to the Initial Charging Time, which is the time required to start charging the vehicle’s battery from its current state [32]. The ICT in Figure 2 is then projected to a weekly pattern for this simulation.

In the weekly EV charging demand simulation, the total charging load at each hour of the week (t = 0, …, 167) is modeled as a convolution of the stochastic initiation of charging sessions ICT with per-vehicle charging power profiles, scaled by the fleet size and charging frequency, and partitioned into AC and DC charging components. Specifically, for each vehicle model i, the fleet size Ni was denoted, and the average weekly session per vehicle

$$f_i={\displaystyle \frac{7}{Charge\_Freq\_day{s}_i}},$$

(1)

and the mode shares s_i,AC and s_i,DC defined by the AC:DC ratio, with

$$s_{i,AC}=ratio\_AC, s_{i,DC}=1-ratio\_AC$$

(2)

when DC capability exists, otherwise s_i,AC = 1. Let p_ICT(h) be the probability of session initiation at week hour h, and let Pi,m($\tau$) for m ∈ {AC, DC} be the discrete power profile (kW) as a function of elapsed charging time $\tau$ (from the parsed times and power curves). The per-model, per-mode demand was calculated via

$$D_{i,m }\left(t\right)=N_i{f}_i{s}_{i,m}{\displaystyle \sum }_{\tau =0}^{T_{i,m}}\mathrm{Ƥ}ICT\left(\left(t-\tau \right)mod 168\right){\mathrm{Ƥ}}_{i,m}\left(\tau \right),$$

(3)

and the overall demand is then

$$D\left(t\right)={\displaystyle \sum }_i{\displaystyle \sum }_{m \in \left\{AC,DC\right\}}{D}_{i,m}\left(t\right),$$

(4)

with the AC and DC subcomponents obtained by restricting the inner sum to m = AC or m = DC. This formulation accurately captures the temporal distribution of charging events via ICT convolution, the heterogeneity of vehicle-specific power curves, and the influence of fleet composition and weekly charging frequency.

2.2. Machine Learning Methods

To forecast long-term grid dynamics in the Jawa–Madura–Bali system, we considered four principal families of time-series models: statistical ARIMA, kernel-based Support Vector Regression (SVR), ensemble Random Forest, and recurrent Long Short-Term Memory (LSTM) networks. Each approach offers distinct advantages and limitations when applied to multivariate, nonlinear hourly load data over multi-decadal horizons.

The Autoregressive Integrated Moving Average (ARIMA) model decomposes the series into autoregressive (AR) terms, differencing operations to enforce stationarity, and moving-average (MA) components. While ARIMA reliably captures linear temporal correlations in well-behaved, stationary series, its performance degrades in the presence of pronounced nonlinearities and exogenous interactions unless extensive feature engineering is undertaken.

Support Vector Regression (SVR) extends the SVM framework to regression by fitting a function within an ε-insensitive tube, leveraging kernel transformations to model nonlinear relationships. When configured on lagged input windows, SVR can mitigate noise and overfitting in high-dimensional settings. However, its reliance on fixed training horizons and the computational burden of kernel and hyperparameter selection limit scalability for long-range forecasting.

Random Forest regression constructs an ensemble of decision trees trained on lagged features and averages their outputs to reduce variance and control overfitting. This nonparametric, ensemble-based approach does not require stationarity assumptions and can capture nonlinear dependencies. Nevertheless, its fixed temporal window restricts memory of distant patterns, hindering the accurate modeling of seasonality and secular trends beyond the training window.

Long Short-Term Memory (LSTM) networks address these limitations by incorporating gated cell states that regulate information flow across time steps. An LSTM cell computes the following operations at each time step t [33]:

$$\begin{array}{c}{i}_t=\sigma \left(x_t{U}^i+h_{t-1}{W}^i\right)\\ f_t=\sigma \left(x_t{U}^f+h_{t-1}{W}^f\right)\\ o_t=\sigma \left(x_t{U}^o+h_{t-1}{W}^o\right)\\ {Ć}_t=\mathit{tanh}\left(x_t{U}^g+h_{t-1}{W}^g\right)\\ C_t=\sigma (f_t\ast C_{t-1}+i_t\ast {Ć}_t)\\ h_t=\mathit{tanh}\left(C_t\ast o_t\right)\end{array}$$

(5)

where:

• x_t is the input vector at time t.
• W and U are weight matrices and bias vectors for the respective gates and candidate cell state.
• σ is the logistic sigmoid activation, and tanh is the hyperbolic tangent activation.
• ⊙ denotes element-wise multiplication.

This architecture prevents gradient vanishing or explosion, enabling the model to learn both short- and long-term dependencies. Given our objective to capture nonlinear interactions over a 36-year horizon, the LSTM network provides the most favorable trade-off between representational capacity and computational tractability. Among various neural network architectures evaluated in the literature, LSTM stands out for its maturity, robustness in time-series forecasting, and relatively lower training complexity. It is particularly well suited for energy systems modeling, where sequential patterns, seasonal variability, and policy-driven shifts must all be captured with temporal coherence. Thus, LSTM was selected as the most appropriate architecture to balance performance with interpretability and scalability in long-horizon EV-grid forecasting.

3. Methods

3.1. Simulation Details

The simulation is implemented in Google Colab using Phyton 3 to leverage its free, cloud-based Python environment, which provides seamless collaboration, scalable compute resources, and straightforward integration with data storage services [34]. This choice ensures reproducibility and ease of access for all collaborators without the need for local software installations.

Simulation Periods:

• ML_TRAIN (1974–2004): Historical data used to fit the machine-learning model parameters.
• ML_VALIDATE (2004–2014): Independent historical subset for hyperparameter tuning and overfitting assessment.
• ML_TEST (2014–2024): Final historical block to evaluate predictive performance prior to forecasting.
• BAU_SIMULATION (2025–2060): Business-as-usual projection of supply and demand under current policy and growth trends.
• ML_CORRECTION (2035–2060): Comparative scenario applying ML-derived adjustments to the BAU baseline to explore alternative outcomes in parallel to BAU Simulation.

These temporal boundaries (y₋₃ through y_Y) are configurable based on specific research and real-world conditions. The years are indexed by an offset y_offset ∈ {−3,−2,−1,0,1,…,X,Y}:

• y−3 corresponds to the first ML training year (1974);
• y−2 marks the first year of ML validation set (2004);
• y−1 marks the first year of ML testing year (2014);
• y0 denotes the last year of ML test-set (2024);
• y1 marks the first year of the BAU projection (2025);
• yX marks the final first stage of BAU (2034);
• yX + 1 through yY cover the ML-corrected projection period (2035–2060).

The simulation begins by training a Long Short-Term Memory (LSTM) network on historical Jamali grid data to capture annual growth patterns in both population-driven demand and PV supply capacity. Using the timeline offsets, the model undergoes sequential training, validation, and testing phases to ensure robust generalization. Once the LSTM has been calibrated, it produces an annual growth projection, namely, the expected percentage increase in EV charging demand and PV generation capacity for each forecast year.

3.2. Supply Demand Simulation

The LSTM simulation’s phase, shown in Figure 3, calculate the annual growth factor. These LSTM-derived growth factors feed directly into the supply–demand balance module, which executes two serial workflows:

• Baseline (BAU) Simulation
• ML-Corrected Projection

As displayed in Figure 4, starting at Year 0 (y0), the model initializes all input variables (e.g., base supply curve, EV population, charging behavior parameters). For each subsequent year y, it computes hourly supply and demand curves, aggregates weekly energy balances, and advances to y + 1 until the end of the BAU horizon yX. Upon completion, the results are plotted to visualize the evolving gap between generation and load.

Under the “business-as-usual” (BAU) assumption, current policy and investment trajectories remain unchanged through the end of the sitting president’s term [35]. Both supply and baseline demand curves continue to grow at their historical AGRs without any EV-related or corrective interventions. Balance Scores are computed for each year to gauge how misalignment evolves if no new policies are adopted.

Beginning at Year yX + 1, the simulation again steps through each year, but now adjusts the baseline supply and demand inputs using LSTM-predicted growth variables. This “corrected” pathway proceeds until yY, at which point the two result sets (BAU vs. ML-corrected) are overlaid for comparative analysis.

All computations and visualizations run in a Google Colab environment—Python 3.11 with TensorFlow 2.x, pandas, and NumPy—leveraging a standard Colab VM (Intel Xeon CPU, ~12 GB RAM). A full multi-decade simulation typically completes in 4–5 min on the default runtime.

3.3. Scoring System for Supply Demand Balance

The alignment between weekly supply S_t and total demand D_t^total is quantified via two complementary metrics:

• Daily Energy Mismatch;
• Shape similarity via Dynamic Time Warping (DTW).

The Daily Energy Mismatch is computed using

$$E_{t,j}^{prod}={\displaystyle \sum }_{h=\left(j-1\right)\times 24+1}^{jx24}{s}_{t,h}, E_{t,j}^{use}={\displaystyle \sum }_{h=\left(j-1\right)\times 24+1}^{jx24}{d}_{t,h}^{total}, j=1,\dots ,7.$$

(6)

then the normalized root-mean-square error (NRMSE) in MWh:

$$NRMS{E}_t={\displaystyle \frac{\sqrt{\frac{1}{7}{\sum }_{j=1}^7{\left(E_{t,j}^{prod}-E_{t,j}^{use}\right)}^2}}{ma{x}_j\max \left(E_{t,j}^{prod}-E_{t,j}^{use}\right)}}\times 100\%,$$

(7)

and define the magnitude score

$$Scor{e}_{mag,t}=100\%-NRMS{E}_t$$

(8)

NRMSE measures daily energy mismatch in MWh across one week and is normalized by the maximum weekly deviation. Expressing this as a percentage (Score mag,t) allows an intuitive interpretation and fair comparison across different years and load levels. Unlike simple power deviation metrics, NRMSE captures aggregated error over time, which is more robust under fluctuations and useful for assessing planning accuracy in long-term simulations.

The Dynamic Time Warping distance between S_t and D_t^total is

$$DT{W}_t=\underset{P\in \mathrm{Ƥ}}{\min }{\displaystyle \sum }_{\left(i,j\right)\in P}\left|S_{t,i}-d_{t,j}^{total}\right|.$$

(9)

We normalize by DTW (S₂₀₂₄, S₂₀₂₄) to obtain

$$Scor{e}_{shape,t}=\left(1-{\displaystyle \frac{DT{W}_t}{DTW\left(S_{2024},S_{2024}\right)}}\right)\times 100\%.$$

(10)

The final balance score for year t is the average:

$$Balance Scor{e}_t={\displaystyle \frac{1}{2}}\left(Scor{e}_{mag,t}+Scor{e}_{shape,t}\right).$$

(11)

DTW, in contrast, captures temporal misalignment between supply and demand. It is sensitive to shifts in peak hours or charging patterns—critical in EV-rich systems. The shape score (Score shape,t) evaluates whether supply follows demand’s structure, not just its total.

The final Balance Score, as an average of both, provides a comprehensive indicator: it penalizes both under/over-supply and temporal mismatches. This dual-metric framework improves decision-making by ensuring both quantity and timing of supply are aligned with EV-driven demand variability. We evaluated both metrics under extreme conditions, confirming the sensitivity of NRMSE to total energy gaps and of DTW to misaligned load curves. These tests validate the Balance Score’s effectiveness in stress conditions.

3.4. Machine Learning and Simulation Accuracy Checks

To assess the reliability of our LSTM-based growth forecasts, we evaluated model performance on the held-out ML_TEST dataset (2014–2024). Four standard error metrics, displayed in

Table 3. Simulation’s output error metrics.

Metric	Value
Root Mean Square Error (MRSE)	0.85%
Mean Absolute Error (MAE)	0.42%
Mean Absolute Percentage Error (MAPE)	1.12%
Coefficient of Determination (R2)	0.94

, were computed as shown in the output graph of the simulation below [36].

4. Simulation Inputs and Outputs

4.1. Machine Learning Data Sets (1974–2004)

Figure 5 shows the entire 1974–2024 series and extend with the 10-year ahead forecast (2025–2034). This view emphasizes the continuity between model validation and long-term projection and yields an average annual growth factor (multiplier) that feeds directly into our next supply–demand simulation stage.

Walk-forward is the de facto standard in time-series forecasting competitions and operational deployments for robust model evaluation. It adapts the model to new patterns by retraining on each new observation, crucial for evolving grid supply dynamics. This method provides a realistic out-of-sample performance estimate to guide supply–demand simulation inputs. The final zoomed plot in Figure 6 underscores how the model leverages recent data for reliable future growth estimation.

Zooming in on 2013–2024, this graph highlights the training window (2013–2017), the in-sample fit period (2018–2024), and the forecast horizon (2025–2034), clearly demarcated by vertical lines. It shows that the post-break LSTM adapts to the accelerated growth regime, providing a precise fit (MAPE ≈ 11%) before projecting forward [36].

The population data shown in Figure 7 are also being considered in the 1974–2024 datasets. The average annual growth factor is 1.58%.

4.2. Beginning of 2024 Data Input

The data input in the beginning of 2024, which is the year0 in the simulation, could be categorized into four segments:

• Initial variable input and its AGR;
• Supply and demand curve in year 0;
• Vehicle data;
• Vehicle charge load.

The initial variable inputs are described in

Table 4. Initial variable input [37,38,39,40,41].

Variable Input	Value in Year 0 (2024)
Supply Capacity in MW	50,103
Demand Peak in MW	42,635
Indonesia Population	281,603,800
Vehicle Ownership Percentage	63.9%
EV Percentage	0.7%
2W:4W Ratio	4.8:5.8
Luxury:Daily Preferences	25:75
Fast:Home Charging Preferences	25:75

. The annual growth rate (AGR) is displayed in

Table 5. Initial annual growth rate.

Variable Input	Initial AGR Value
Default Supply Curve AGR	1.05
Default Demand Curve AGR	1.05
Population AGR	1.011
Vehicle Ownership Rate AGR	1.036
EV Penetration Rate AGR	1.05
2W Ratio AGR	1.01
Daily Ratio AGR	1.02
AC Charging Ratio AGR	1.03

.

The baseline supply and demand curves for each simulation year were derived from peer-reviewed data and our prior MDPI publication, then adjusted using PLN Statistics for 2024 to ensure consistency with actual grid performance [37,42,43,44]. The simulation assumes a single aggregated supply capacity variable that evolves over time based on historical data; it does not disaggregate generation types or dispatch behavior. This allows the focus to remain on evaluating demand-side management strategies under constrained supply conditions. These weekly profiles represent one typical week of the Jawa–Madura–Bali grid’s supply–demand balance, reflecting diurnal and weekday–weekend variation without seasonal bias, given the region’s equatorial climate. Weekends induce discernible troughs in both supply and demand, as illustrated in Figure 8a,b. In the initial year (year 0), the combined supply and demand profile (Figure 8c) demonstrates lower power levels during weekend hours; this composite curve forms the reference state from which subsequent annual growth and corrective scenarios evolve.

The weekly ICT profile, shown in Figure 9, represents the probability that an EV begins charging at each hour of a typical week. This distribution was derived from national driving-pattern surveys [32] and reflects higher charging start probabilities during end of the day evening, with reduced likelihood during work hours and weekends.

Below is a summary of the eight electric-vehicle models included in our simulation, spanning both passenger cars and two-wheelers. For each vehicle we specify its nominal driving range (200–620 km), battery capacity (1.44–103 kWh), DC-fast-charging capability (where available), typical charge duration for both DC and AC modes, peak charging power, assumed days between charges (30 days), and the year production began. These parameters form the basis of our charging-load profiles and directly feed into the demand-side modelling [45,46,47,48,49,50,51,52,53,54].

Table 6. Vehicle specification [45,46,47,48,49,50,51,52,53,54].

Vehicle Type	Var in Simulation	2W/4W	Daily/Luxury	Driving Range (km)	Battery Size (kWh)	DC Fast Charging	DC Charging Duration (min)	DC Max Charging Load (kW)	AC Charging Duration (min)	AC Max Charging Load (kW)	Daily Commute (km) [47]	Charge Frequency (days)	Sales in 2025 (Jan-Mar) [46]	Market Share	Yearly Sales in y0	Production Start	Price in 2025	References
BYD Denza D9 EV	EV4W_L_1_F/S	4W	Luxury	620	103	Yes	30	166	600	11	30	20.6	2524	55%	18,039	2022	950,000,000	[48]
Chery iCar 03 (J6)	EV4W_L_2_F/S	4W	Luxury	401	50.6	Yes	30	60	480	7.7	30	13.3	2034	45%	14,537	2023	505,500,000	[49]
Wuling Air EV	EV4W_D_1_F/S	4W	Daily	200	17.3	No	0	0	510	6.6	30	6.6	1111	27%	26,625	2022	184,000,000	[50]
BYD M6	EV4W_D_2_F/S	4W	Daily	420	55.4	Yes	70	115	780	7	30	14	2967	73%	71,104	2024	383,000,000	[51]
Gesits G1	EV2W_L_1	2W	Luxury	50	1.44	No	0	0	240	0.3	30	1.6	1908	87%	136,480	2019	372,700,00	[52]
United TX1800	EV2W_L_2	2W	Luxury	60	1.5	No	0	0	300	1.3	30	2	278	13%	19,886	2022	359,000,000	[53]
Viar NQ1	EV2W_D_1	2W	Daily	60	1.5	No	0	0	300	0.24	30	2	416	37%	175,806	2022	220,000,000	[54]
Polytron Fox R	EV2W_D_2	2W	Daily	130	3.7	No	0	0	270	0.82	30	4.3	694	63%	293,292	2024	205,000,000	[55]

below lists the EV models included in our simulation, chosen for their market popularity and representative features. These eight models were selected based on their significant market presence in Indonesia, diversity in battery capacities and powertrain configurations, as well as data availability on technical specifications. Importantly, the selection spans both two- and four-wheeled EVs, ensuring a balanced representation of the most relevant vehicle types currently influencing national load profiles. We apply a price-based classification to distinguish daily-use from luxury vehicles: any four-wheeler priced below IDR 400 million or two-wheeler priced below IDR 30 million is treated as a daily-oriented model in this study, while those above these thresholds are considered luxury segment examples [55]. This categorization informs scenario definitions and ICT parameter assignments.

Figure 10 illustrates the charging power profiles used as inputs for this simulation, distinguishing between slow (AC) and fast (DC) charging modes according to a prescribed ratio—vehicles without DC capability are modeled as charging at AC 100% throughout. The upper panels plot instantaneous charging power (y-axis) against state of charge (SOC, x-axis) for a representative set of vehicles: the left shows AC curves peaking between 0.5 and 11 kW, the right DC curves reaching up to 170 kW. These generic SOC-based charge curves—adopted in lieu of proprietary data and without accounting for battery-health effects—provide a reasonable first-order approximation of real behavior. The lower panels reframe the same profiles over charging duration (hours on the x-axis): the left for AC, the right for DC. Here, the pronounced variation in power over time highlights how load intermittency is driven by vehicle class and by the timing of charging initiation—key factors in assessing the temporal clustering of EV demand on the grid.

4.3. End of 2024 Simulation Output

Once all input parameters are specified, the year 0 simulation calculates the baseline EV population for each model. Using initial market share estimates, sales forecasts for 2025 (January–March), and the regional total vehicle stock, the algorithm allocates the number of units per model. Figure 11 presents a snapshot of the simulation output, displaying the computed vehicle counts by model that form the starting fleet for subsequent charging and load analyses.

Following initialization, the simulation calculates EV charging demand for each model over the reference week. Figure 12a presents the per-model demand results, while Figure 12b differentiates between AC and DC charging profiles, applying the power-duration curves described earlier. These graphs quantify hourly load contributions from each vehicle type, combining ICT start probabilities with model-specific charging rates to form the aggregate EV demand input for the first simulation year.

The aggregated EV demand profile is subsequently superimposed on the baseline system demand, comprising residential, commercial, and industrial loads, to produce the total load curve for year 0, displayed in Figure 13.

The hourly balance score is defined as the percentage ratio of supplied energy to demanded energy at a given hour. The weekly average balance score is then obtained by averaging the 168 individual hourly scores. Figure 14 illustrates both the hourly balance profile and its weekly mean, highlighting periods of surplus and deficit that inform corrective-action decisions.

To complement the weekly analysis, we compare daily supply and demand energies to reveal intraday mismatches. Figure 15 presents the total 24-h supply–demand profile for the reference week in year 0.

4.4. 2029 BAU Output

After establishing the initial conditions, we simulate a business-as-usual (BAU) scenario over the next five years, aligning with the national presidential term length. This five-year buffer accounts for the lead time required to translate measurement insights into policy and infrastructure changes. As shown in Figure 16, the grid balance remains robust throughout this period, indicating that current growth rates do not immediately compromise reliability.

Figure 17 presents the corresponding daily energy supply–demand profiles. Both visuals confirm that, even without corrective measures, the system sustains adequate performance over the first term.

Figure 18 further demonstrates that the weekly average balance score steadily increases over the five-year BAU simulation horizon. This trend occurs because supply and demand trajectories become more closely aligned over time, reflecting the effect of consistent growth rates and the scaling of generation capacity in line with rising consumption.

4.5. 2060 BAU Output

Extending the BAU simulation through to year 2060 reveals a pronounced divergence between supply and demand trajectories. Figure 19 illustrates the weekly balance comparison for year 0 and year 2060, highlighting substantial deficits during peak hours. The corresponding daily energy profiles (Figure 20) underscore severe shortfalls in the evening charging window, driven by the aggregate growth of the EV fleet.

Figure 21 depicts the projected population growth of each EV model under the BAU scenario, with separate curves for every vehicle type from the initial year through 2060. These trajectories reflect model-specific adoption rates and illustrate how the expanding fleet contributes to the escalating charging demand.

4.6. 2060 Optimized Output

The simulation then applies the corrective actions recommended for 2035 to project grid performance through to 2060. The selection of 2035 as the earliest feasible intervention point reflects both technical and political realities in Indonesia. Given that national EV development is highly dependent on political leadership, particularly presidential policy direction, and with presidential elections occurring every five years, the next leadership transition in 2029 represents a key turning point. Assuming a conservative policy uptake trajectory, 2035 marks the latest plausible year after a new president’s first term to enact meaningful energy reform. This ensures that even under slow policy adoption, the proposed interventions remain politically and practically viable.

Table 7. Optimized AGR.

Variable Input	Initial AGR Value	Optimized AGR	AGR Difference
Default Supply Curve AGR	1.05	1.09108	0.0410799
Default Demand Curve AGR	1.05	1.09356	0.0435602
Population AGR	1.011	1.011	0
Vehicle Ownership Rate AGR	1.036	1.036	0
EV Penetration Rate AGR	1.05	1	−0.05
2W Ratio AGR	1.01	0.984036	−0.0259642
Daily Ratio AGR	1.02	1.06369	0.04369
AC Charging Ratio AGR	1.03	0.996017	−0.0339827

summarizes the initial annual growth rate (AGR) settings alongside the optimized AGR values suggested by the model—derived by minimizing projected supply–demand deficits—and the percentage difference between them. These adjustments indicate the precise scaling of demand and supply growth to implement in 2035, yielding the best system balance by 2060. By comparing BAU and optimized AGRs, stakeholders can identify which parameters require tightening or relaxation to prevent the large deficits observed in the unconstrained scenario.

If these corrective measures are fully implemented in 2035, the annual balance score under the optimized scenario steadily surpasses that of BAU. Figure 22 illustrates that by 2038 the optimized trajectory yields a higher average weekly balance, marking the first divergence where BAU underperforms. Furthermore, Figure 23 demonstrates that by 2042 the daily supply–demand gap in the BAU simulation expands sharply, indicating a critical loss of equilibrium. In contrast, the optimized scenario maintains manageable deficits, confirming the long-term efficacy of the 2035 interventions in preserving grid stability.

Figure 24 presents the stark contrast in weekly balance scores for the year 2060 between the BAU and optimized scenarios. In the zoomed inset of this diagram, the model’s effort to tightly align supply and demand is evident, particularly around hour 125 of the week, where the optimized curve closely tracks the supply profile, minimizing deviations that would otherwise appear under the BAU trajectory.

Figure 25 presents the aggregate 24-h energy provided and consumed for year 2060 under both BAU and optimized scenarios. In the zoomed inset of Figure 25 the daily energy difference in the optimized scenario remains minimal.

This annualized energy difference translates into significant cost savings. Beginning in 2035 and extending through 2060, the optimized scenario contributes cumulative savings of approximately IDR 16,435 billion, as illustrated in Figure 26 (PLN production cost of Rp1.300/kWh [56]).

Figure 27 further demonstrates that the simulation can track the growth rate of each EV model under optimized AGRs. From 2035 onward, individual vehicle trajectories deviate from the BAU path, reflecting the corrective AGR adjustments.

5. Discussion

The simulation outcomes for the 2060 planning horizon indicate that imposing a constraint on EV charging demand growth—ensuring its annual growth rate remains at or below the supply AGR from 2035 onward—yields substantial system-wide benefits. Under this corrective action, cumulative energy savings reach approximately 12 TWh by 2060, corresponding to avoided generation costs of roughly IDR 16,435 billion (based on an assumed average production cost of IDR 1300/kWh).

Deferral of these measures until 2047 reduces energy savings to 5 TWh, lowers cost savings to IDR 3500 billion, and degrades the grid’s average hourly balance score by 8 percentage points. These findings highlight the compounding advantage of early intervention: delays exacerbate supply–demand mismatches and necessitate more aggressive, capital-intensive capacity expansions in later years. Achieving the required alignment by 2035 involves coordinated deployment of demand-side management, vehicle preference incentives, and adaptive capacity planning of EV population.

However, translating these modeled interventions into real-world implementation poses practical challenges. These include uncertainties in future EV adoption rates, policy enforcement delays, variability in consumer behavior, and uneven development of charging infrastructure across regions. Institutional coordination across ministries and utility stakeholders, data availability for localized calibration, and budgetary constraints in expanding grid capacity are also critical barriers. Addressing these obstacles requires phased regulatory support, targeted public–private investment schemes, and continuous monitoring of implementation outcomes.

While the model is calibrated to the Java–Madura–Bali (Jamali) grid, the simulation framework is designed to be adaptable to other regions within Indonesia and to developing countries with similar grid and transport dynamics. Its flexibility lies in the modularity of input parameters, including supply infrastructure, EV adoption profiles, spatial density, and policy targets, which can be reconfigured to match local conditions. However, boundary constraints such as area-specific infrastructure saturation and population-driven energy intensity must be considered when transferring the model to new geographies.

6. Conclusions

This study presents a comprehensive simulation framework for the Jawa–Madura–Bali grid that integrates vehicle charging demand growth, business as usual growth rate and policy interventions through 2060. To support timely implementation of recommended interventions, several targeted regulatory actions are needed. These include fiscal incentives aligned with grid readiness (simulated AGR as parameter growth) and integration of demand-side management into national energy planning. These policies must be aligned across sectors and updated regularly to reflect evolving EV adoption and supply dynamics.

The simplification of modeling charging as a full 0–100 percent state-of-charge cycle enables clear quantification of aggregate demand impacts; future work should refine this assumption by incorporating more accurate represent partial-charge behavior. The framework omits explicit representation of charging infrastructure costs and spatial availability, excludes V2G impacts on battery degradation, and assumes homogeneous driving and charging patterns across the region. Addressing these limitations in subsequent research—by integrating infrastructure investment models, spatially resolved travel behavior, stochastic renewable generation, and targeted regulatory scenarios—will further enhance the realism and policy relevance of the approach.

Ultimately, this work underscores the critical role of early, integrated policy alignment in delivering energy, economic, and reliability gains during the transition to a decarbonized transport sector.