Overcoming the Cold-Start Problem in EV Charging Forecasting
The rapid growth of electric vehicle (EV) adoption has been imposing pressure on distribution systems [1]. Voltage drop and congestion of feeders are the most common issues at the distribution networks level resulting from EV chargers. Smart charging is a well-accepted strategy to address these challenges [2]. However, a practical smart charging strategy requires a reasonably accurate representation of EV charging behavior. In particular, smart charging systems benefit from estimates of the required energy and amount of time available to charge the EV, as two essential parameters of charging behavior [3].
Deep neural networks (DNNs) have been widely applied to EV behavior forecasting. DNNs need a very large number of samples for training, without which the models are prone to over and under-fitting. For newly committed EV owners, however, not much data is initially available. Even a year’s worth of charging records leads to only 365 charging samples, which is not adequate to train a deep forecasting model. Hence, forecasting the EV charging behavior of a newly committed EV suffers from the so-called cold-start forecast problem, i.e., there is insufficient historical data to train the model [4].
This blog post suggests a transfer learning-based framework using a deep generative model to address the cold-start forecasting of newly committed EV charging behavior. The solution discussed in this blog targets two variables of interest that are required for smart charging: plug-out time and required energy per charging event. This blog post is based on one of the publication of Grid Foresight Lab [5]. For more information, please visit this link.
The main idea is to enable using DNNs for EV charging behaviour forecasting of recently committed EV owners. Instead of training a DNN from scratch for each new user, a source model is first trained on charging events from available EV charging behaviour datasets. This source model learns general EV charging behavioral patterns that are shared across households rather than being specific to any individual. By capturing this shared pattern, the source model learns a prior knowledge that can be used to forecast the same behavioral patterns for a new EV owner.
Hence, the knowledge is then transferred to a target model built for the new user. It should be noted that in this domain, knowledge refers to the weights of DNNs. Therefore, when training a model to forecast the charging behaviour of a new user, rather than initializing weights randomly, the target model begins with the pretrained weights from the source model. This initialization provides a meaningful starting point, allowing the model to produce reasonable forecasts even before adequate user-specific data has been fed to the model.
Despite the aforementioned strategy, it is worth mentioning that although transfer learning might mitigate the cold-start problem, it does not fully resolve it when the available user-specific dataset is extremely small. Data augmentation is another practical complementary strategy in addition to transfer learning to address the lack of adequate training data. Accordingly, a generative adversarial network is suggested to augment the training set of the new EV owner. This specific type of generative model produces tabular synthetic charging samples conditioned on the distributional characteristics of the new user’s early records. The synthetic samples are expected to improve the robustness of the training procedure and reduce overfitting.
In addition to the suggested strategies above, another fundamental choice in training DNNs is how pretrained weights are used during fine-tuning. Rather than allowing all layers of the deep model to update, we suggest freezing all neurons located in the middle layers of the target DNNs and making a shortcut between the output and input layers during the backpropagation procedure. As a matter of fact, freezing a layer means that the weights of that layer will not change during the fine-tuning step, with the aim of preserving the feature extractor of the base learner (source model). These layers remain in the DNN for the feedforward phase. Not only does this approach decrease the computational cost but also avoids so-called catastrophic forgetting, i.e., the tendency of a DNN to forget previously learned knowledge.
From a practical point of view, the suggested solution in this blog could inform utilities about the stalled period of EVs, which is indeed the available period to implement smart charging on participating customers in demand response program. A utility can activate smart charging for a new EV owner as soon as a vehicle is connected to the grid, rather than waiting until a large number of charging events have been accumulated. This matters because the early adoption phase is precisely when unmanaged charging is most disruptive to distribution feeders.
Want to Learn More?
For full details, including methodology nuances and results, check out the papers of Gridforesight team published in IEEE Transactions on Intelligent Vehicles journal.
If you would like to connect about applying these methods to your case or want to discuss future research, reach out at ali.forootani1@ucalgary.ca.
Reference
[1] X. Wang, Y. Chen, L. Du and Y. Chen, “Online Deadline-Aware EV Charging Considering State-of-Charge-Dependent Peak Charging Rate,” in IEEE Transactions on Smart Grid, vol. 16, no. 6, pp. 5275-5288, Nov. 2025, doi: 10.1109/TSG.2025.3591027.
[2] G. Ruan and M. A. Dahleh, “Temperature-Controlled Smart Charging for Electric Vehicles in Cold Climates,” in IEEE Transactions on Smart Grid, vol. 16, no. 3, pp. 2197-2207, May 2025, doi: 10.1109/TSG.2025.3525672.
[3] O. Frendo, N. Gaertner, and H. Stuckenschmidt, “Improving smart charging prioritization by predicting electric vehicle departure time,” IEEE Trans. Intell. Transp. Syst., vol. 22, no. 10, pp. 6646–6653, Oct. 2021.
[4] M. Zhang, W. Li, E. Wang and J. Wu, “MapT-STC: A One-Shot Spatiotemporal Data Completion Framework for Cold-Start Tasks in Sparse Crowdsensing,” in IEEE Transactions on Mobile Computing, doi: 10.1109/TMC.2026.3678980.
[5] A. Forootani, M. Rastegar, and H. Zareipour, “Transfer learning-based framework enhanced by deep generative model for cold-start forecasting of residential ev charging behavior,” IEEE Transactions on Intelligent Vehicles, vol. 9, no. 1, pp. 190–198, 2023.
