Multi-scale fusion transformer for EV charging station load prediction

Why smarter EV charging matters

As electric vehicles (EVs) spread through cities, their charging habits are starting to matter as much to the power grid as rush hour traffic does to highways. If many drivers plug in at once, local cables and transformers can be stressed; if grid operators know in advance when and where charging will spike, they can balance supply, avoid blackouts and even use EVs as flexible energy resources. This paper introduces a new way to forecast how busy EV charging stations will be over the next one to four days, aiming to give grid planners and charging companies a much clearer view of the road ahead.

The challenge of guessing tomorrow’s plugs

Predicting EV charging demand is surprisingly hard. Drivers choose when to charge based on work schedules, weather, errands and even traffic jams, making the total load at a station jump up and down in a seemingly random way. Traditional statistical tools, which work well for smoother patterns, struggle with these sharp swings and with the mix of influences from outside the power system. Even modern deep-learning approaches, such as recurrent neural networks and standard Transformer models, often capture either long-term trends or short-term jitters, but not both at once, and they usually treat outside factors like weather and traffic in a fairly crude way.

A new model that sees time in layers

To tackle these issues, the authors design a Multi-scale Fusion Transformer (MFT), a machine-learning model tailored to EV charging stations. At its core is a "multi-scale" mechanism that lets the model look at past charging data with several different lenses at the same time. One lens focuses on broad, slow changes over days; another zooms in on rapid hour‑to‑hour swings; others lie in between. By guiding separate attention heads within the model to specialize in different time scales, and then fusing their views, MFT can track both the overall rhythm of charging and the fine details of sudden peaks and dips more effectively than a standard Transformer.

Teaching the model what weather and traffic really mean

Charging demand does not depend on time alone. Using data from a large residential area in Norway, the researchers add 14 outside features to the model, including temperatures, wind, rainfall, sunshine and how many vehicles are moving through several nearby zones each hour. First, they perform a broad statistical scan to see how strongly each factor is linked to charging over the whole dataset. Sunshine, for example, shows a noticeable negative relationship: brighter days often mean fewer charging events there. This analysis produces a set of base importance weights that tell the model, in general terms, which factors tend to matter more and which matter less.

Letting the model adapt from hour to hour

Of course, any single day can break the average pattern: sometimes the weather is calm but traffic is chaotic, or vice versa. To adapt to these shifts, the MFT includes a multi‑variable fusion module that adjusts the feature weights for each prediction. It uses a cross‑attention step in which the current charging pattern "queries" the external data, deciding which weather or traffic signals are most relevant right now. The model then blends these signals into a compact representation of the outside world, which it combines with the multi‑scale view of past loads. A downstream decoder turns this joint picture into forecasts for the next 24, 48, 72 or 96 hours.

How well the new approach performs

The team trains and tests MFT using real hourly data, comparing it with well‑known forecasting models such as GRU, LSTM, bidirectional LSTM and a standard Transformer. Across all prediction horizons and error measures, the new model comes out ahead, and its edge widens as the forecast stretches farther into the future. On average, MFT reduces key error metrics by more than 20 percent compared with recurrent networks and by around 10 percent compared with a plain Transformer. Importantly, it maintains stable accuracy even for 72‑ and 96‑hour forecasts, whereas other models tend to drift and lag behind actual load swings.

What this means for everyday energy use

For non‑specialists, the takeaway is that better mathematics can make EV charging quietly more reliable and efficient. By combining a layered view of time with a flexible understanding of weather and traffic, the Multi-scale Fusion Transformer offers grid operators a sharper forecast of how heavily stations will be used in the coming days. That, in turn, can support smarter scheduling of power plants, smoother integration of renewable energy and more informed placement of new chargers. As EV adoption grows and future models add battery behavior into the mix, tools like MFT could become key ingredients in keeping electric transport both convenient for drivers and friendly to the grid.

Citation: Liu, W., Qiao, J., Wang, W. et al. Multi-scale fusion transformer for EV charging station load prediction. Sci Rep 16, 8609 (2026). https://doi.org/10.1038/s41598-026-38562-z

Keywords: electric vehicle charging, load forecasting, deep learning, transformer model, smart grid