Smart Cities Gain Edge with New EV Charging Pricing Strategy
As urban centers around the world race to embrace sustainable mobility, a groundbreaking study has unveiled a novel approach to electric vehicle (EV) charging that could redefine how cities manage energy and transportation systems. The research, led by Lu Siyue and a team of engineers from State Grid Beijing Electric Power Company, introduces an innovative pricing model designed to harmonize renewable energy integration with urban traffic flow. Published in the September 2024 issue of Computer Applications and Software, this work presents a dual-layer optimization framework that not only minimizes societal costs but also enhances grid stability and reduces congestion in metropolitan areas.
The study addresses one of the most pressing challenges in modern urban planning: the growing interdependence between power distribution networks (DN) and urban transportation networks (UTN). As EV adoption accelerates, uncoordinated charging behavior can lead to localized grid overloads, inefficient use of renewable resources, and increased traffic congestion at charging stations. Traditional pricing models often fail to account for these systemic interactions, focusing instead on isolated aspects such as time-of-use tariffs or static location-based fees. However, Lu Siyue’s team recognized that a more holistic approach was necessary—one that considers both the dynamic nature of wind power generation and the stochastic patterns of travel demand.
At the heart of their methodology is a two-tiered optimization model. The upper layer focuses on determining optimal charging service fees (CSF) within a distribution network that incorporates wind power. This layer is formulated as a second-order cone programming (SOCP) problem, which allows for precise modeling of electrical constraints while accounting for the inherent uncertainty in wind output. By treating CSF as a decision variable, the model enables grid operators to influence EV charging behavior in real time, steering load away from periods of low renewable availability or high network stress.
The lower layer of the model simulates user behavior through a traffic assignment problem based on the principle of user equilibrium (UE). Unlike traditional models that assume deterministic travel patterns, this framework incorporates randomness in origin-destination (OD) traffic flows, reflecting the unpredictable nature of daily commutes and errands. Drivers are assumed to make rational choices, selecting routes and charging stops that minimize their individual costs—including travel time, waiting time, and electricity expenses. The interaction between these two layers creates a feedback loop where pricing signals shape traffic patterns, which in turn affect power demand and grid conditions.
What sets this research apart is its integration of deep reinforcement learning (DRL) to solve the resulting stochastic bilevel problem. Conventional optimization techniques struggle with such complex, coupled systems, especially when uncertainty is involved. DRL, however, offers a powerful alternative by leveraging neural networks to approximate optimal policies through trial and error. The team implemented both gradient-based and gradient-free training algorithms, comparing their performance in terms of convergence speed and solution quality. Notably, the gradient-free deep genetic strategy (DGP) demonstrated superior performance, achieving results closer to the theoretical optimum than its gradient-based counterpart, deep deterministic policy gradient (DDPG).
The practical implications of this work are significant. In a five-node test system, the application of optimized CSF reduced average social cost by 14.9%, from 10,022.6 yuan to 8,543.2 yuan. More importantly, wind energy utilization improved dramatically, with curtailment rates dropping from 31.1% to just 4.8%. This means that nearly all available wind power was successfully absorbed into the grid, reducing waste and enhancing sustainability. Traffic flow also became more balanced across the network, alleviating bottlenecks on heavily used corridors while increasing throughput on underutilized routes. Although individual driver costs rose slightly—from 1.43 yuan to 1.45 yuan per trip—the overall societal benefit was substantial, demonstrating that minor trade-offs at the individual level can yield major gains for the collective.
To validate scalability, the researchers applied their framework to a real-world urban scenario involving 39 nodes and 10 EV charging stations. The city-scale network included diverse road types—from ring roads to inner-city arterials—and was powered by a distribution grid with eight distributed generators and three wind farms. Under uncertain OD demand and fluctuating wind output, the DRL-driven pricing strategy achieved a 2.05% reduction in total system cost, bringing it down from 35,158.0 yuan to 34,436.9 yuan. Wind energy consumption rates increased significantly at key nodes: from 21.3% to 62.6% at node 15, and from 31.6% to 95.3% at node 33. These improvements were accompanied by measurable reductions in traffic congestion, particularly on high-capacity links that had previously approached saturation.
One of the most compelling aspects of the study is its emphasis on out-of-sample performance. Real-world conditions rarely match historical data perfectly, so any pricing strategy must be robust to unforeseen variations. The team tested their model under altered probability distributions for both traffic demand and wind generation, simulating scenarios where average OD volumes shifted from 1,000 to 1,100 vehicles and wind output dropped from 10 MW to 8 MW. Despite these changes, the system maintained strong performance, indicating that the learned policy generalizes well beyond the training environment. This resilience is crucial for real-world deployment, where weather patterns and travel behaviors are inherently volatile.
The research also highlights the limitations of non-deep reinforcement learning methods in large-scale applications. When compared against a fitted Q-iteration (FQI) algorithm with a discretized action space, the DRL approach outperformed significantly. While FQI managed to converge for some charging stations, it failed to identify optimal pricing levels across the entire network, ultimately achieving only a 0.20% cost reduction—less than one-tenth of what the DRL method accomplished. This stark contrast underscores the importance of continuous action spaces and high-dimensional state representations in managing complex, interconnected systems.
From a policy perspective, the findings suggest that centralized coordination between utility operators and transportation authorities can unlock substantial efficiencies. In many cities, these domains operate independently, leading to suboptimal outcomes. For example, a utility might offer lower electricity prices during off-peak hours without considering whether drivers are willing or able to charge at that time. Similarly, traffic planners may expand road capacity without accounting for the spatial distribution of charging infrastructure. The proposed framework bridges this gap by enabling joint optimization of energy and mobility resources.
Moreover, the model supports a shift from reactive to proactive management. Instead of waiting for congestion or grid instability to occur, operators can anticipate problems and adjust pricing signals accordingly. For instance, if a forecast predicts high wind generation during the afternoon, CSF can be lowered at nearby charging stations to encourage EV owners to charge during that window. Conversely, if traffic sensors detect rising congestion near a popular charging hub, fees can be increased temporarily to divert drivers to less crowded alternatives. This dynamic responsiveness enhances system flexibility and improves user experience.
The role of artificial intelligence in this transformation cannot be overstated. While earlier attempts at smart charging relied on rule-based systems or simple optimization routines, the integration of deep learning enables adaptive, data-driven decision-making. The neural network learns complex relationships between variables—such as how a 10% increase in wind output affects optimal pricing at different locations—and applies this knowledge in real time. Over successive iterations, the policy becomes increasingly refined, approaching near-optimal performance even in highly stochastic environments.
Another advantage of the DRL approach is its ability to handle partial observability. In practice, operators do not have perfect information about every vehicle’s state of charge, destination, or preferred route. The model accounts for this uncertainty by treating observed data—such as aggregate traffic flow and wind speed—as sufficient statistics for making decisions. Through repeated interaction with the environment, the agent learns to infer hidden states and act accordingly, much like a human dispatcher would.
The implications extend beyond EV charging. The same framework could be applied to other shared infrastructure systems, such as public transit, bike-sharing, or micro-mobility services. By treating pricing as a control variable and using machine learning to optimize it, cities can achieve better resource allocation, reduce environmental impact, and improve quality of life. As urban populations continue to grow, such tools will become essential for maintaining livability and sustainability.
Critically, the study adheres to principles of transparency and reproducibility. All simulations were conducted using open-source solvers (Mosek and Baron) within a Python environment, and the DRL framework was built using PyTorch, a widely adopted deep learning library. The inclusion of benchmark comparisons—against both scenario-based stochastic programming and non-deep RL methods—ensures that claims about performance are rigorously substantiated. Furthermore, the use of real-world network topologies and realistic parameter values enhances external validity, making the results more applicable to actual urban settings.
Ethical considerations are also addressed implicitly through the focus on social cost minimization. Rather than maximizing profit for charging station operators or minimizing cost for individual users, the objective function prioritizes collective welfare. This aligns with the public service mission of utility companies and municipal governments, ensuring that technological advancements serve the broader community. Additionally, the slight increase in individual costs observed in the five-node system suggests that equity concerns were balanced against efficiency gains, avoiding scenarios where only a subset of users bears the burden of system optimization.
In conclusion, the work by Lu Siyue and colleagues represents a significant step forward in the integration of energy and transportation systems. By combining rigorous mathematical modeling with cutting-edge machine learning techniques, they have developed a pricing strategy that is both technically sound and practically viable. The results demonstrate clear benefits in terms of cost reduction, renewable energy utilization, and traffic management. As cities worldwide seek to decarbonize their transport sectors and modernize their grids, this research provides a blueprint for intelligent, coordinated infrastructure operation. With further refinement and field testing, such models could soon become standard tools in the urban planner’s toolkit, helping to build smarter, greener, and more resilient cities.
Lu Siyue, Ji Hongquan, Zhang Lu, Xu Hui, Wang Peiyi, State Grid Beijing Electric Power Company, Computer Applications and Software, DOI: 10.3969/j.issn.1000-386x.2024.09.053