The rapid proliferation of electric vehicles (EVs) necessitates efficient charging infrastructure management to ensure grid stability and optimal resource utilization. This paper proposes a machine learning (ML)-based model that predicts stochastic EV charging demand while explicitly incorporating the effects of atmospheric temperature on charging behavior. By leveraging historical data and time-series analysis, the model accurately captures spatiotemporal dependencies to forecast demand with high precision, enabling proactive capacity planning. To demonstrate practical utility, we apply the predicted demand to optimize station capacity using a data-driven ML framework. Compared to existing optimization strategies, our approach achieves 18% greater capacity efficiency and reduces required charging station resources by 11.06% per EV class while maintaining low Loss of Load Probability (LoLP). These improvements result from the model’s ability to anticipate arrivals 1-2 hours in advance, enabling dynamic resource allocation. The framework also facilitates real-time decision-making for admitting or rejecting charging requests, effectively preventing station overload.