1. Introduction
With the global shift toward renewable energy and the collective commitment to carbon neutrality, the installation of distributed renewable generators has been accelerating at an unprecedented rate. According to data from the International Renewable Energy Agency [
1], 585 GW of renewable energy capacity was added worldwide in 2024, with China accounting for approximately 63.8% of this total. At the distribution grid level, projections suggest that by the end of 2025, China could deploy over 500 GW of renewable generators and more than 120 million charging infrastructure units to support its 2060 carbon neutrality target [
2]. This transition from fossil fuels to cleaner, distributed energy sources presents both opportunities and challenges for the dynamic operation of power distribution networks. Meanwhile, the inherent uncertainty and variability of renewable energy, particularly widely adopted photovoltaic (PV) systems, pose significant challenges to voltage regulation, power loss reduction, and overall system efficiency. Consequently, traditional passive distribution networks are being reconfigured into active distribution networks (ADNs), which incorporate real-time monitoring [
3], intelligent dispatch [
4], and flexible control enabled by advanced power electronic interfaces and distributed energy resources (DERs).
In this context, enhancing operational efficiency has become a critical objective for the secure and economic operation of ADNs. Efficient operation not only ensures the optimal utilization of renewable resources and controllable assets such as on-load tap changers (OLTCs) and capacitor banks (CBs) but also supports the network in maintaining fine regulation under uncertain operating conditions [
5]. To address the operational challenges introduced by stochastic-generated renewable energy while maintaining system efficiency, power-electronics-based technologies have been increasingly explored in recent years. Among them, soft open points (SOPs) and electric springs (ESs) have emerged as two promising solutions for enhancing the operational efficiency of ADNs.
A SOP is a power electronic device that replaces traditional mechanical tie-switches at normally open points in distribution networks [
6], as shown in
Figure 1. Both terminals of SOP are voltage source converters (VSCs), and are decoupled by the DC capacitance. Owing to the DC isolation, each terminal of SOP can be controlled independently, thus enabling fast and continuous control of active and reactive power flow between two terminals, improving voltage profiles, reducing network losses, and facilitating higher DER penetration. Preliminary studies have investigated the application of SOPs in distribution system operation. For instance, ref. [
7] integrates SOPs to enhance voltage stability and reduce losses, while the author of [
8] employs SOP for imbalance mitigation in ADNs. In [
9], a coordinated control strategy between SOPs and energy storage systems is proposed to improve operational flexibility. Under normal operating conditions, SOPs have demonstrated effectiveness in network reconfiguration and optimal power flow management [
10]. Furthermore, stochastic and robust optimization techniques have been applied to optimize SOP placement and operation under renewable generation uncertainty [
11,
12,
13].
ES is an advanced power electronics apparatus deployed on the demand side, installed between a distribution bus and non-critical loads (NLs). The main structure of ES is a voltage source inverter (VSI), making it both cost-effective and highly responsive in voltage and reactive power control, as shown in
Figure 2. Unlike conventional compensators, the ES exclusively regulates reactive power, eliminating the need for costly active power converters [
14]. When combined with NLs, the series structure forms a smart load (SL), operating in parallel with critical loads (CLs) to enhance grid efficiency [
15]. In ADN with ESs, NLs refer to electric loads that can undertake relatively wide voltage and power variations. On the other hand, CLs require a stable power and voltage supply. By adjusting the voltages of ESs and NLs, ESs can provide real-time voltage support and absorb surplus renewable energy. This makes ES particularly suitable for applications in systems with high penetration of distributed PV units. The concept was first introduced in [
16], where the fundamental operating principle and circuit topology were established. Since then, various control strategies have been developed to improve ES performance. Droop and distributed control methods are proposed in [
17,
18] for voltage regulation in low-voltage networks with high PV integration. In [
19], ESs are utilized to improve ADNs’ efficiency through loss reduction and voltage flattening. Additional benefits of ESs include enhanced system flexibility [
20], resilience [
21], imbalance mitigation [
22], and frequency support [
23].
While both SOPs and ESs contribute to voltage regulation and operational efficiency, they operate at different levels of the distribution system: SOPs act at the network level by controlling inter-feeder power flows, whereas ESs function at the demand side by modulating end-user load consumption. Existing studies typically treat SOPs and ESs in isolation, with limited focus on their coordinated operation across multiple timescales under stochastic renewable generation. Therefore, to bridge this gap, this paper proposes a two-stage stochastic coordination framework that jointly dispatches SOPs and ESs to enhance ADN operational efficiency while accounting for uncertainties. The key contributions are summarized as follows:
A two-stage stochastic programming model is proposed to enhance operational efficiency in ADNs with hourly and intra-hour timescale coordination control. The model accounts for uncertainties in PV output and load demand to improve solution robustness. The hourly-stage control is realized by operating legacy mechanical devices (OLTC and CB) as “here-and-now” first-stage actions. The intra-hour stages are realized by joint dispatch of SOPs and ESs to act as “wait-and-see” second-stage actions to compensate for the first-stage solutions.
A computationally efficient solution methodology that convexifies non-convex power flow and SOP capacity constraints, while employing a data–knowledge hybrid-driven approach for ES modeling is proposed. A multi-layer perceptron (MLP) learns the nonlinear ES operational characteristics and is subsequently linearized using mixed-integer programming. Finally, the optimization problem is reformulated into a tractable mixed-integer second-order cone programming model (MISOCP) that can be handled reliably and efficiently with commercial solvers.
Numerical simulations are conducted based on the IEEE 33-bus distribution system, revealing that the proposed two-stage stochastic coordination strategy effectively enhances the operational efficiency of ADNs. The hybrid solution methodology reduces computational burden while maintaining high solution accuracy.
The rest of the paper is organized as follows:
Section 2 gives the problem formulation, organized into a two-stage stochastic programming model.
Section 3 details the solution methodology. The case study in
Section 4 validates the proposed approach. Lastly,
Section 5 finalizes this paper with conclusions.
4. Case Study and Analysis
The case study is conducted on the modified IEEE 33-bus test system, as illustrated in
Figure 5a. An OLTC is connected at the substation node with a 10% voltage regulation range. A CB with 300 kVar reactive power compensation capability is installed at bus 25. Five 700 kW PV plants are integrated into the network. Two SOPs, each rated at 500 kVA (with 300 kVar reactive power capability), are placed between buses 11 and 21. Eight ESs are distributed across the system. NLs and CLs are assumed in a 70:30 ratio [
30], with NLs tolerating up to 10% voltage variation as per (
40). Note that the load is predefined and remains constant, regardless of voltage changes. The objective function weights are set to 0.833 and 0.167 for the respective terms. The prediction errors of PV and load demand follow the Gaussian distribution with standard deviations of 10% and 5%, respectively. A total of 1000 Monte Carlo scenarios are generated and reduced to 20 representative scenarios using a distance-based reduction method.
Table 1 presents the remaining parameters.
Figure 5b shows the 15-minute profiles of the PV output and load demand in the test system. The nominal power of the load is predefined and remains constant irrespective of voltage fluctuations. Consequently, the actual load is determined by multiplying the nominal power by the load ratios provided in
Figure 5b.
The training dataset for the ES voltage constraints is generated using the full ES model. Specifically, voltage pairs
are randomly sampled within the bounds defined by the NL voltage constraint (
24) and the system voltage limit (
26). These voltage pairs, along with the load impedance angle
, are input to the ES voltage equation (
21) to compute the corresponding reactive power output
. Over 10,000 feasible data samples are collected for training and testing the ES MLP model. During generation, any sample violating the ES capacity constraint (
25) is discarded to ensure feasibility.
The optimization model is gathered with CVXPY 1.7 and solved with Gurobi 12.0. The regression MLP is constructed using PyTorch 2.7.1 with a batch size of 128 and 600 epochs. All numerical experiments are conducted on an Intel (R) i5-10210U 1.6 GHz CPU with 16 GB of memory.
4.1. Regulation Performance
First, the regulation performance of the proposed two-stage stochastic coordination strategy is evaluated by comparing it with two benchmark cases: (1) an ADN without any regulation schemes and (2) an ADN employing only hourly dispatch of OLTC and CBs.
Figure 6 illustrates the voltage profiles across the test system under these different cases, where red and blue dashed lines refer to the safe and flattening limits, respectively.
As shown in
Figure 6a, in the absence of regulation, bus voltages violate both the safe upper limit during periods of high PV generation and the lower limit during peak demand periods. This highlights significant voltage security challenges in unregulated ADN operations. When hourly dispatch of OLTC and CB is applied, as in
Figure 6b, voltage violations—particularly at the lower (safe and flattening) bounds—are significantly mitigated, showing high voltage quality. This improvement arises because the OLTC maintains a higher secondary-side voltage during peak load periods, supporting voltage levels across the network. However, during peak generation periods, due to the intermittent and stochastic nature of PV generation within intra-hour intervals, the fixed-hourly OLTC and CB settings cannot adapt dynamically, limiting their ability to suppress overvoltages during rapid PV output fluctuations. Moreover, the constant reactive power injection from CBs can lead to overcompensation during periods of reduced load or high PV output, resulting in intra-hour overvoltage violations. It is worth noting that control of OLTC and CB creates a foundational platform for the second-stage operation of SOPs and ESs. Notably, in the absence of regulation, there are 21 instances of 15 min violations of 56 violations of flattening limits. With Stage 1 alone, these violations are reduced to 41, representing a reduction of 26.8%. This improvement underscored the critical contribution of Stage 1 to voltage flattening.
Figure 6c shows the voltage profiles of the test system based on the proposed approach, demonstrating that the two-stage stochastic coordination framework effectively leverages SOPs and ESs for intra-hour regulation on a 15 min timescale, complementing the hourly OLTC/CB actions. Under this strategy, all bus voltages remain strictly within the prescribed safety limits—neither exceeding the upper bound nor falling below the lower bound. Furthermore, the secondary objective
, which promotes voltage flattening, ensures that all node voltages are maintained within ±3% of the nominal value. This tight regulation reflects a high degree of voltage quality control and demonstrates the strategy’s capability to achieve not only voltage security but also superior voltage stability and quality.
Figure 7 presents a comparison of power losses in the test network under different operational scenarios. As shown, power loss reaches its peak during high-demand periods due to heavy loading on the distribution lines. Similarly, during periods of peak PV generation, reverse power flow can also lead to increased losses. In the absence of any regulation, power losses are the highest across all time intervals, exceeding those observed under both the hourly regulation scheme and the proposed two-stage strategy. Both during peak demand and peak generation periods, the proposed strategy achieves the lowest power loss. Moreover, throughout the remaining operating periods, it consistently maintains lower losses compared to the other cases.
As shown in
Figure 7b, in the unregulated case, the maximum, minimum, and average power losses reach 193.1 kW, 26.3 kW, and 71.3 kW, respectively, resulting in a total power loss of 1710.4 kWh over 24 h. Such high losses significantly reduce system efficiency, leading to higher operational costs. With hourly regulation by OLTC and CB, losses are reduced to 164.0 kW (maximum), 18.4 kW (minimum), and 59.8 kW (average), yielding a total loss of 1435.2 kWh, which is a reduction of 16.1% compared to the unregulated case. This demonstrates that hourly control can effectively help to mitigate losses, not only improving operational efficiency but also setting the stage for further loss minimization in Stage 2, which highlights the synergistic effect between the two stages and greatly contributes to the ADN operation.
However, the proposed two-stage stochastic coordination strategy further reduces the maximum, minimum, and average power losses to 104.5 kW, 17.3 kW, and 41.0 kW, respectively. The total power loss is reduced to 983.5 kWh, representing a 42.5% reduction compared to the unregulated case and a 31% reduction compared to hourly regulation. This substantial improvement highlights the superior operational efficiency enabled by the proposed strategy.
4.2. Comparative Analysis
Then, we conduct a comparative analysis to demonstrate the advantages of two-stage stochastic coordination over a single-stage strategy, and another comparative analysis to demonstrate the benefit of joint dispatch over independent dispatch of SOP or ES.
4.2.1. Comparison with Single-Stage Strategy
The proposed strategy in (
49) is a two-stage stochastic programming model that explicitly accounts for intra-hour uncertainties in PV output and load demand. It leverages the fast response capabilities of ESs and SOP to mitigate rapid fluctuations within each hour. In contrast, the single-stage strategy neglects the expectation term
in (
49), thereby ignoring intra-hour uncertainties in the decision-making process. As a result, the dispatch of OLTC, CB, SOP, and ES remains fixed throughout the hour, and ESs and SOPs are not dynamically adjusted in real time.
Figure 8 illustrates the voltage profiles of the test system under 200 random scenarios (generated via Monte Carlo simulation) from 13:00 to 14:00, where red and blue dashed lines refer to the safe and flattening limits, respectively. Since the single-stage strategy does not account for potential intra-hour variations in load and PV generation, it adopts a riskier operating point: the secondary-side voltage of the OLTC is set to 1.01 p.u., and the CB provides a fixed reactive power compensation of 300 kVar. Consequently, the maximum bus voltages approach the voltage flattening upper bound of 1.03 p.u. Without adaptive adjustment from ESs and SOPs, approximately 8.5% of the scenarios exceed the safety limit (1.05 p.u.), and around 45% violate the flattening limit (1.03 p.u.), demonstrating that the single-stage approach fails to handle intra-hour variability due to its neglect of uncertainty.
In stark contrast, the proposed two-stage stochastic coordination strategy incorporates second-stage uncertainty realizations into the first-stage decisions, resulting in a more conservative and robust solution. It sets the OLTC secondary voltage to 1.00 p.u. and the CB compensation to 200 kVar, thereby reducing the initial voltage stress. Under random scenarios, ESs and SOP adaptively respond to intra-hour fluctuations based on real-time monitoring, ensuring that all bus voltages remain within both the safety bounds (0.95–1.05 p.u.) and the tighter flattening bounds (0.97–1.03 p.u.).
It can be concluded that the single-stage strategy fails to fully exploit the fast response characteristics of power electronic devices and does not consider intra-hour uncertainties, leading to significant voltage violations during the hour. In contrast, the proposed two-stage stochastic programming model integrates the potential realizations of second-stage uncertainties, significantly enhancing regulation performance and operational robustness. The hourly dispatch of OLTC and CB establishes a robust foundation for the intra-hour control actions conducted by SOPs and ESs in Stage 2. This conservative first-stage solution significantly supports the intra-hour adjustments, ensuring that uncertainties are effectively managed. By enabling ESs and SOPs to rapidly adapt to intra-hour variations, their dynamic response capabilities are fully utilized, greatly contributing to operational efficiency enhancement in ADNs.
4.2.2. Benefit Analysis on Joint Dispatch of SOP and ES
The benefit brought by the joint dispatch of SOP and ES on enhancing the operational efficiency of ADN is analyzed by comparing it with two different schemes. The hourly dispatches are all conducted by the OLTC and CB, while the following schemes realize the second-stage (intra-hour) dispatches:
- Scheme A:
Intra-hour regulation solely based on SOP.
- Scheme B:
Intra-hour regulation solely based on ESs.
- Scheme C:
Intra-hour regulation based on joint dispatch of SOP and ESs (proposed).
Scheme A and Scheme B correspond to the optimization problem in (
49) without ESs and without SOP, respectively. In contrast, the proposed Scheme C utilizes the full model, incorporating both SOP and ESs. The voltage violation ratio is computed as the total duration of violations divided by the overall simulation time.
As shown in
Table 2, relying solely on hourly regulation via OLTC and CB results in voltage violations exceeding the safety limit for 9.3% of the time, posing significant risks to system security. Although Scheme A reduces these violations to 6.2%, it still fails to eliminate them entirely and cannot meet the stricter voltage flattening requirement, which remains violated for 12.5% of the time. In contrast, Scheme B, utilizing eight distributed ESs, completely eliminates both safety and flattening limit violations. This superior performance can be attributed to the spatial flexibility and rapid response capabilities of distributed ESs, which offer greater regulation flexibility compared to a single SOP. Then, with the joint dispatch of eight ESs and SOP (proposed Scheme C), voltage regulation performance remains excellent, with zero voltage violations under both criteria and all bus voltages kept within the desired flattening bounds at all times.
Regarding network power losses, Scheme A and Scheme B achieve reductions of 10.95% and 18.12%, respectively, compared to the hourly-only scheme. In contrast, the proposed joint dispatch reduces losses by 31.47%, significantly outperforming both Scheme A and Scheme B. This indicates that coordinating ESs and SOP yields lower power loss than relying solely on either device alone.
Compared to Scheme A, which solely depends on SOP for intra-hour regulation, the proposed scheme considers ESs for cooperation, thus realizing no voltage violations and lower losses. Compared to Scheme B, which solely depends on ESs for intra-hour regulation, the proposed scheme incorporates SOP for cooperation, thus achieving even lower power loss. Overall, the proposed joint dispatch scheme outperforms both Scheme A and Scheme B. The joint dispatch of SOP and ES not only ensures operational safety but also significantly enhances the overall operational efficiency of ADNs. It demonstrates superior performance in both voltage regulation and power loss reduction compared to using SOP or ES independently.
In conclusion, the above comparative analyses demonstrate that the proposed two-stage stochastic coordination strategy significantly improves intra-hour voltage regulation, ensures operational robustness, and minimizes power losses while maintaining zero voltage violations. By leveraging the fast response capabilities of SOP and ESs to complement the slower OLTC and CB actions, the strategy effectively bridges temporal gaps in control. This synergistic coordination paves the way for more reliable and efficient management of ADNs.
4.3. Algorithm Validation
Further, this section gives validation for the proposed solution algorithm from two perspectives: the performance of the trained MLP in characterizing ES and the computational efficiency of the reformulated MISOCP model.
The MLP is trained offline using collected operation data and is used to approximate the nonlinear ES voltage constraints. In this study, we employ a shallow MLP architecture consisting of 3 neurons in the input layer, 15 neurons in a single hidden layer, and 1 neuron in the output layer. Training is performed using the stochastic gradient descent (SGD) optimizer on a standard CPU. The average training time is approximately 50 s, which is negligible in the context of offline model development.
Figure 9 illustrates the training loss and prediction accuracy of the MLP. As shown, the training loss decreases rapidly and converges to a small value, indicating effective learning.
Figure 9b presents the prediction errors for both capacitive and inductive operating modes of the ES. The maximum prediction errors are below 0.8 kVar, while the average errors do not exceed 0.2 kVar for either mode. These results demonstrate that the trained MLP accurately captures the nonlinear operational characteristics of ESs, enabling high-fidelity representation within the optimization framework. Given the high accuracy achieved with such a lightweight network, the overall complexity of the data-driven MLP is very low.
By replicating the ES nonlinear constraint with the well-trained MLP and applying convexification techniques to the power flow and SOP constraints, the original MINLP problem is reformulated as a tractable MISOCP. This reformulated model is solved using the commercial solver Gurobi. For comparison, the original MINLP is solved using two approaches: (1) the BARON solver, which employs a duality-based branch-and-bound algorithm for the MINLP model; (2) Gurobi’s built-in spatial branching method with outer approximation, which is a global optimization method [
31].
MINLP is generally NP-hard and requires exponential time in the worst case. The global optima of MINLP cannot be guaranteed. As shown in
Table 3, due to the presence of multiple scenarios, inherent nonlinearity, and non-convexity, BARON fails to converge and returns an infeasible solution, highlighting the computational difficulty of the original MINLP formulation. Gurobi’s spatial branching method successfully obtains a solution within 275 s, providing a globally optimal approximation of MINLP.
In contrast, even though the MISOCP problem is still NP-hard in general, it is known to be polynomially solvable in the case of convex problems if all nonlinear constraints are second-order cone representable. Based on the modern branch-and-bound method and accelerating methods, the MISOCP empirically has better scalability than the MINLP solution. The structured convexity enables MISOCP frameworks to grow polynomially with problem size, while MINLP often exhibits exponential growth in runtime due to the challenges of exploring nonlinear and non-convex solution spaces. As shown in
Table 3, the proposed MISOCP approach achieves convergence in only 101.2 s, with an average deviation of just 0.1% from the reference global objective value obtained by Gurobi. This represents a reduction in computation time of over 60%, while maintaining extremely high solution accuracy. These results demonstrate that the proposed methodology significantly alleviates the computational burden and ensures robust convergence, making it highly effective for solving the two-stage stochastic coordination strategy in ADNs. The computational convergence also demonstrates the lower complexity of MISOCP over MINLP, improving tractability compared to solving the original MINLP directly.
Overall, the proposed MISOCP solution with linearized MLP demonstrates significantly reduced computational complexity, allowing the use of off-the-shelf solvers with improved tractability compared to solving the original MINLP directly.