A Dual-Tower Model for Station-Level Electric Vehicle Charging Demand Prediction

With the rapid growth of electric vehicles (EVs), the efficient operation of EV charging stations has become crucial. Accurate demand prediction enables operators to respond effectively to demand fluctuations. This paper proposes a novel dual-tower architecture to predict demand changes caused by both price-related and price-unrelated factors. The left tower uses a Bi-GRU hybrid with 1D convolutional layers to predict baseline demand from price-unrelated factors. The right tower constructs a heterogeneous graph of stations to model price elasticity. An elasticity function then calculates the actual demand based on price elasticity and baseline charging. Our model employed meta-learning in the pre-training step to transfer knowledge from region-level to station-level data for enhancing accuracy. Experimenting on a real-world dataset from Zeekr Intelligent Technology with data from over 800 stations across China, our model significantly outperforms two comparison methods and could predict the charging demand effectively.