Abstract: Intermittent spare parts are typically characterized by high value, demand intermittency, and critical importance. Accurate demand forecasting is essential for optimizing inventory management, ensuring timely maintenance, and maintaining the reliability and safety required for normal operation. To address this issue, an adaptive Q-learning-based demand forecasting method for intermittent spare parts is proposed. First, similarity metrics and K-means clustering are used to preprocess and cluster the demand data. Then, a Q-learning algorithm is adopted to dynamically adjust the matching relationship between the spare parts clusters and forecasting models, enabling adaptive model selection to improve forecasting accuracy. Finally, the effectiveness of the proposed method is validated using a dataset of intermittent spare parts. Results show that the adaptive Q-learning can significantly reduce the forecasting errors for spare parts, with three evaluation metrics E_nrmse, E_mae, and E_mase all being lower than those using six benchmark models, including Croston, ARIMA, and Deep Renewal Exact. The average values across the three metrics are 8.921 0, 1.485 6, and 0.395 0, respectively. These findings demonstrate that the proposed adaptive Q-learning method provides accurate and reliable forecasting results and is well-suited for predicting the demand of intermittent spare parts.