A multi-dimensional feature aggregation network for electric vehicle charging demand prediction

Why smarter charging forecasts matter

As electric cars spread through our cities, the simple act of plugging in is becoming a complex puzzle. Power companies and city planners must make sure chargers are in the right places, the grid can handle the load, and prices encourage drivers to plug in at the best times. This paper introduces a new way to forecast when and where drivers will need to charge, aiming to make those forecasts not only more accurate but also faster and cheaper to compute.

The challenge of guessing tomorrow’s plugs

Predicting electric vehicle charging demand sounds straightforward: look at past usage and project it forward. In reality, it is tangled in many overlapping patterns. Stations in nearby neighborhoods influence one another because drivers can choose between them. Charging demand rises and falls through the day and week in regular rhythms. On top of that, outside factors such as electricity prices, weather, holidays, and traffic all nudge drivers’ choices. Many recent forecasting methods focus mainly on geography and time, but treat these extra influences one by one, missing how they interact together and over time. Others bring in detailed “physics-informed” rules about how prices and demand should behave, which can improve accuracy but often require heavy computation and detailed prior knowledge.

A new way to blend many signals

The authors propose a model called the Multi-Dimensional Feature Aggregation Network, or MDFANet, built specifically for urban charging demand. Rather than feeding raw data straight into a complex spatiotemporal model, MDFANet first passes the data through a dedicated “feature enhancement” stage. In this stage, the model looks both along the time axis and across the different variables—such as demand, price, and other features—to distill a compact “feature core” that captures how these signals move together. This core is then blended back with the original inputs, producing a richer but still lightweight description of the data that is easier for later components to learn from.

How the model learns cities in motion

Once the features are enhanced, MDFANet applies a two-part spatiotemporal module. First, a graph-based attention network focuses on relationships between zones in the city, assigning more weight to areas whose charging behavior is most informative for each target zone. Multiple attention “heads” help the model consider several types of spatial influence at once, while a special residual link prevents important details from being washed out as layers are stacked. Next, a recurrent time-series module models how these spatially informed features change over time, learning short- and longer-term patterns in charging behavior. Finally, a simple decoder translates the learned internal representation into forecasts of future charging demand at different lead times, from 15 minutes up to an hour ahead.

Putting the method to the test

The researchers evaluated MDFANet on a public dataset from Shenzhen, China, which records charging demand and real-time prices every five minutes across 247 zones. They compared their model against a variety of approaches, including classic statistical tools, standard deep learning networks, graph-based models, and the latest physics-informed benchmark. Across four common accuracy measures, MDFANet consistently achieved lower forecasting errors than all competitors, including the physics-informed model PIAST. At the same time, MDFANet required roughly half the training time of these physics-informed methods under the same hardware and training schedule. Tests that deliberately removed parts of the model showed that both the feature aggregation module and the spatiotemporal module were necessary to reach the best performance. Additional experiments that reduced the amount of training data suggested that MDFANet degrades gracefully, maintaining useful accuracy even when only half of the original data are available.

What this means for drivers and cities

In simple terms, this work shows that we can get better and faster forecasts of when and where electric vehicles will need to charge by first carefully blending all the different signals in the data, and then learning how locations and times interact. For city planners and grid operators, more accurate and efficient predictions could support smarter placement of charging stations, smoother grid operation, and more responsive pricing schemes. For drivers, it could translate into shorter waits and more reliable access to chargers. The study still relies mainly on past charging and price records, and the authors point out that future versions should incorporate richer context such as weather and traffic. Even so, their results suggest that thoughtfully designed data-driven models like MDFANet can help cities keep pace with the rapid growth of electric vehicles.

Citation: Yu, Y., He, L., Yu, Z. et al. A multi-dimensional feature aggregation network for electric vehicle charging demand prediction. Sci Rep 16, 13181 (2026). https://doi.org/10.1038/s41598-026-38855-3

Keywords: electric vehicle charging, demand forecasting, spatiotemporal modeling, deep learning, smart grid