Ammar Kamoona

Ammar Kamoona

AI/ML Researcher | RMIT Research Fellow | Powercor AI Scientist

Detecting EV charging events online with a memory-augmented transformer

1 minute read

Published:

Distribution networks in Australia are seeing more electric vehicles every year, and network operators need to know, in near real time, when and where charging events happen. Smart-meter data offers the raw signal, but EV charging looks a lot like other high-power appliances (heat pumps, kettles, ovens) once the meter aggregates everything at the household level.

In our paper in Applied Energy (2025) we treat this as an online event detection problem on a time series. The core idea is a memory-augmented transformer that keeps a compact bank of past patterns and attends over them as new smart-meter samples arrive. The memory gives the model a longer effective context than the raw input window, and the attention lets it match the current signal against a small set of “prototype” charging shapes learned during training.

Two design choices matter in practice. First, the model runs in a streaming setting, so decisions have to be made without seeing the end of a session. We train with a causal mask and evaluate on rolling windows to match this. Second, we keep the memory small so the model can run inside a utility’s edge or gateway stack, not just in a research notebook.

The method sits behind part of the NexusCharge project with ABB Australia and the ARC Linkage electrification program with AGL, AusNet, and SpendWatt.

Reference. Kamoona A.M., Lazarevic L., Al Khafaf N., Ali S.M.N., Jalili M., Razzaghi R. Online electric vehicle charging detection based on a memory-augmented transformer architecture. Applied Energy, 377, 124549, 2025.

You May Also Enjoy

Future Blog Post

less than 1 minute read

Published:

This post will show up by default. To disable scheduling of future posts, edit config.yml and set future: false.

Temporal co-learning for residential electricity forecasting

less than 1 minute read

Published:

Residential electricity forecasting sits between two hard problems. Household-level load is noisy and non-stationary, while feeder-level load is smoother but loses the behavioural signal that drives the peaks. Models trained on one level of aggregation often generalise badly to another.

What growing EV adoption does to residential load profiles

1 minute read

Published:

Australian distribution networks are planning for an era in which a large share of households charge an EV at home. The interesting question is not just “how much extra energy?” but “when does that energy get drawn, and how does the shape of the residential load change?”

Point patterns, energy, and defect detection: an RFS view of anomalies

1 minute read

Published:

Defect detection on manufactured parts is a classic anomaly detection problem: most items are normal, defects are rare, and the visual signature of a defect varies from batch to batch. A common practical trick is to extract a set of local features from an image and score how “unusual” the set looks compared with clean examples.