PIN diodes are critical devices for preventing damage from strong electromagnetic signals. Accurately predicting the temperature rise curve of the PIN diode has important guiding significance for selecting protective devices. Machine learning-based methods can effectively predict the characteristics of devices. However, the temperature rise characteristic curve of the PIN diode contains strong nonlinearity and small fluctuations, and traditional machine learning methods cannot predict accurately. To accurately predict the temperature rise characteristic curve of PIN diodes, this paper proposes a prediction method that combines variational mode decomposition (VMD) and autoencoder to decompose the temperature rise characteristics into sub-signals, which include high-frequency fluctuations, intermediate quantities, and low-frequency trend quantities. Then an autoencoder is used to predict each component. Finally, the predicted values of the components are added together, so as to accurately predict the temperature rise characteristic curve of the PIN diode. By comparing with various machine learning methods, it is verified that combining VMD decomposition can effectively improve prediction accuracy, and the advantage of autoencoder in fitting characteristic curves is also verified.