With the rapid growth in electric vehicle ownership, the stochasticity and volatility of charging loads have posed significant challenges to the secure and stable operation of power grids. Therefore, accurate EV charging load forecasting has become a critical requirement for the development of new-type power systems. Existing studies often overlook the heterogeneity of service objects among different charging piles, resulting in the mixing of charging load characteristics associated with different vehicle usage types and limiting the ability of forecasting results to support grid risk assessment and decision-making. To address this issue, a probabilistic EV charging load forecasting method based on K-means clustering and long short-term memory networks (LSTM) is proposed. First, K-means clustering is employed to classify charging piles into five clusters according to their load patterns and dominant service vehicle types, namely private vehicles, buses, official vehicles, taxis, and other vehicles, thereby mitigating the feature-mixing effect. Subsequently, a LSTM-based forecasting model is constructed, in which a weighted quantile loss function is adopted instead of the conventional mean squared error loss to generate forecasting results within the 10%~90% quantile prediction interval. To validate the effectiveness of the proposed model, experiments are conducted using charging pile load data provided by a provincial power grid company. The results show that, after charging pile clustering, the average mean squared error of load forecasting is reduced by 16.6%. The proposed method can provide effective support for risk assessment and dispatch decision-making in power grid operation.