In the aftermarket service of core manufacturing enterprises, accurate demand forecasting of spare parts is crucial for safety inventory decision. However, the intermittent characteristic of work order data hinders existing methods from capturing demand evolutional patterns, thus reducing forecasting reliability. Focusing on the seasonal fluctuation of continuous non-zero demands, this paper proposes an adaptive forecasting method for intermittent time series based on temporal patch similarity. First, a novel Intermittent Predictability Measurement Index (IPMI) is developed to quantify the predictability across three dimensions: noise level, periodic stability, and short-term dependency. Second, a trend-seasonal decoupled long-short term adaptive forecasting model is established. By partitioning the original sequence into different temporal patches, the model utilizes deep convolutional networks and multi-head attention to extract intra-patch short-term features and inter-patch long-term trends, respectively. An IPMI-weighted loss function is further constructed to enable adaptive learning of patch similarity. Finally, kernel density estimation is applied to calibrate the results, mitigating the impact of historical extreme demand. Experimental results on the M5 competition dataset and real-world rail transit spare parts data from China demonstrate that the proposed method accurately predicts the occurrence and intensity of intermittent demand, and also significantly enhances the rationality of safety inventory decision under demand volatility.