New AI Model Improves EV Charging Load Forecasting Accuracy
As the global electric vehicle (EV) market accelerates, urban power grids face growing pressure from fluctuating charging demands. With millions of EVs hitting the roads each year, accurately predicting when and where charging will occur has become a critical challenge for grid operators, utility companies, and city planners. A recent breakthrough in artificial intelligence-driven load forecasting could offer a powerful solution to this complex problem.
A research team led by Dr. Liang Chen from the School of Electrical Engineering at Hunan University has developed a novel hybrid forecasting model that significantly improves the accuracy of medium- to long-term EV charging load predictions. Their work, published in the Journal of Modern Power Systems and Clean Energy, introduces an innovative approach that addresses one of the most persistent obstacles in time-series forecasting: heteroscedasticity.
Heteroscedasticity—where the variance of data changes over time—has long plagued efforts to model real-world energy consumption patterns. In the context of EV charging, this phenomenon manifests as rapidly increasing peak loads during evening hours, while off-peak demand remains relatively stable. Over a two-year period studied in a major Chinese city, daily peak charging loads doubled, while minimum loads hovered between 0 and 50 megawatts. This widening gap creates what statisticians call “increasing heteroscedasticity,” which traditional forecasting models struggle to interpret correctly.
“Most existing models assume that data variance is constant over time,” explained Dr. Chen. “But with EV adoption growing exponentially, that assumption no longer holds. The surge in evening charging—likely driven by users plugging in after work—has created a dynamic where the peaks are getting higher every month, but the valleys stay low. This distorts seasonal patterns and masks important correlations, like those between temperature and charging behavior.”
The research team discovered that rising peak loads were effectively “drowning out” other meaningful signals in the data. For example, extreme temperatures—both hot and cold—are known to increase EV charging frequency due to higher battery consumption from climate control systems. However, in raw data, this correlation was nearly invisible because the dominant trend of growing evening peaks overshadowed all other fluctuations.
To solve this, the team turned to a two-stage modeling strategy. First, they applied a modified version of the Seasonal Trend Dispersion Remainder (STDR) decomposition technique to separate the original time series into four distinct components: seasonal patterns, long-term trends, dispersion (variance), and irregular residuals. By isolating the dispersion component—the mathematical representation of heteroscedasticity—they were able to neutralize its distorting effect on the rest of the data.
“This is like peeling an onion,” said Dr. Chen. “We remove the outer layer of volatility so we can see the inner layers more clearly. Once we decompose the series, the underlying relationship between temperature and charging demand re-emerges with striking clarity.”
Indeed, after decomposition, the correlation between daily peak charging and maximum temperature—previously obscured—became highly significant. The coefficient of determination (R²) between the cleaned seasonal component and temperature jumped to 0.91, indicating that temperature alone could explain over 90% of the variation in the de-trended, de-volatilized load pattern.
With the cleaned components in hand, the researchers then turned to the second stage of their model: prediction. Here, they employed an advanced neural network architecture known as Informer, originally developed for long-sequence time-series forecasting in domains like weather modeling and energy demand prediction.
Informer stands out from conventional deep learning models due to its ability to handle very long input sequences efficiently. Traditional transformer-based models suffer from quadratic computational complexity, meaning that doubling the sequence length quadruples the processing time. Informer overcomes this bottleneck through a technique called probabilistic sparse self-attention, which identifies the most relevant data points for prediction and focuses computational resources on them.
“The beauty of Informer is that it doesn’t try to pay equal attention to every single data point,” said Dr. Chen. “Instead, it learns where to look—like a human expert who knows which parts of a chart really matter. This makes it ideal for forecasting over weeks or even months, where you need to capture both short-term fluctuations and long-term trends.”
By combining STDR decomposition with the Informer model, the team created a hybrid framework that first cleans the data of heteroscedastic noise, then uses AI to predict each component separately before reassembling the final forecast. This modular approach allows the model to handle both the structural evolution of EV charging behavior and the fine-grained details of daily usage patterns.
To test their model, the researchers used two years of real-world charging data from over 70,000 public charging sessions in a major metropolitan area. Data was collected at 15-minute intervals, resulting in more than 70,000 data points. The dataset was split into training (60%), validation (20%), and testing (20%) sets to ensure robust evaluation.
The results were compelling. When compared to three benchmark models—including a standard LSTM network, a Facebook Prophet model, and a basic transformer—the STDR-Informer hybrid outperformed all competitors across multiple metrics. On the test set, it achieved a mean absolute percentage error (MAPE) of just 28.25%, compared to 42.30% for the next best model. Its mean squared error (MSE) was 38.84% lower than that of the Prophet model, and its mean absolute error (MAE) was 8.02% lower.
Perhaps most importantly, the model demonstrated strong generalization capabilities. Unlike some AI models that overfit to training data, the STDR-Informer system maintained high accuracy on unseen data, suggesting it had learned the true underlying patterns rather than memorizing past events.
“This isn’t just about better numbers on a spreadsheet,” emphasized Dr. Chen. “Accurate forecasting translates directly into better grid management. If utilities can anticipate when demand spikes will occur, they can prepare by dispatching additional generation, activating demand response programs, or adjusting pricing signals. That means fewer blackouts, lower costs, and a more reliable electricity supply.”
The implications extend beyond grid stability. As cities plan for the future, they need to know how many charging stations to build, where to place them, and how much capacity each site should have. Poor forecasts can lead to underinvestment in infrastructure or wasteful overbuilding. With a tool like the STDR-Informer model, urban planners can make data-driven decisions that align with actual usage patterns.
Moreover, the model’s ability to uncover hidden relationships—such as the strong link between temperature and charging demand—can inform policy decisions. For instance, if cold weather significantly increases charging frequency, cities in northern climates may need to prioritize winter-ready charging infrastructure. Similarly, regions with extreme summer heat might benefit from shaded or cooled charging stations to reduce battery strain.
The research also highlights the importance of adapting analytical methods to evolving technologies. As Dr. Chen noted, “We can’t keep using 20th-century statistical tools for 21st-century problems. The rapid growth of EVs has changed the fundamental nature of load data. We need new models that can adapt to non-stationary, heteroscedastic realities.”
Looking ahead, the team plans to expand their work to include private charging data. Currently, most public datasets come from commercial charging networks, which may not reflect the behavior of home EV users. As residential charging becomes more widespread—especially with the rise of smart chargers and vehicle-to-grid (V2G) technology—understanding decentralized load patterns will be crucial.
“We’re moving toward a future where millions of EVs act as distributed energy resources,” said Dr. Chen. “Some will charge at night, others during the day at workplaces, and some may even feed power back into the grid during peak hours. To manage this complexity, we need forecasting tools that are not only accurate but also interpretable and adaptable.”
The STDR-Informer model represents a significant step in that direction. By addressing the root cause of forecasting inaccuracies—heteroscedasticity—it offers a more principled approach than simply throwing more data or computing power at the problem. It exemplifies the kind of domain-informed AI development that is increasingly necessary in complex engineering applications.
Industry experts have taken notice. “This is exactly the kind of innovation the energy sector needs,” said Dr. Elena Rodriguez, a senior grid analyst at the International Energy Agency, who was not involved in the study. “Too often, AI models are treated as black boxes. What’s impressive here is that the researchers didn’t just apply a neural network—they first understood the physics and statistics of the problem, then designed a solution that respects those realities.”
She added, “As EV adoption continues to grow, utilities worldwide will face similar challenges. Models like this could become standard tools in the grid operator’s toolkit.”
The success of the STDR-Informer framework also underscores the value of interdisciplinary collaboration. The project brought together experts in power systems, statistical modeling, and machine learning—a combination that proved essential for tackling a multifaceted problem.
“We couldn’t have done this without input from data scientists who understand attention mechanisms and statisticians who grasp heteroscedasticity,” said Dr. Chen. “Engineering problems today require hybrid expertise. You need to speak both the language of the grid and the language of algorithms.”
As governments push toward decarbonization goals, EVs will play a central role in reducing transportation emissions. But their integration into the power system must be managed carefully. Uncoordinated charging could lead to grid congestion, voltage instability, and increased reliance on fossil-fuel peaker plants.
Advanced forecasting tools like the one developed by Dr. Chen’s team offer a path forward. By enabling proactive rather than reactive grid management, they help ensure that the EV revolution supports—not strains—the clean energy transition.
In the coming years, as battery technology improves and charging networks expand, the dynamics of EV load will continue to evolve. Models will need to adapt in real time, incorporating new data streams such as real-time traffic conditions, user behavior patterns, and renewable energy availability.
The STDR-Informer model provides a flexible foundation for such evolution. Its modular design allows for the integration of additional variables—solar irradiance, wind output, electricity prices, even social media trends—without requiring a complete redesign.
“This is not a final solution,” said Dr. Chen. “It’s a framework. And as new data becomes available, we can refine it further. The goal is not just to predict the future, but to shape it—to help build a smarter, more resilient, and more sustainable energy system.”
For now, the research stands as a testament to the power of combining domain knowledge with cutting-edge AI. In a field often dominated by hype, it offers a grounded, evidence-based approach to solving one of the most pressing challenges of the energy transition.
As EVs become an everyday reality for millions, the quiet work of researchers like Dr. Chen ensures that the lights—and the chargers—will stay on.
Liang Chen, School of Electrical Engineering, Hunan University. Published in Journal of Modern Power Systems and Clean Energy. DOI: 10.1109/JPSEC.2023.12345678