In many DPSs, the energy distribution and transmission for each heat source is closely related to its adjacent neighbors, and the thermal dynamics have obvious time-series characteristics. To modeling these processes, a spatiotemporal modeling method is developed in combination with the advantages of deformation CNN (D-CNN) and LSTM in feature extraction and time-series modeling. First, the improved D-CNN strategy is used to extract the spatial features of the DPS. By using the deformable convolution kernel in the convolution layer, the convolution kernel weight and the local sequence are convolved to obtain the preliminary feature matrix, and the multi-scale context information of the space is captured to extract the main spatial features. Then, a cascade LSTM model is constructed to model the time-series dynamics. The sliding window method is used to take the sequential data at several consecutive sampling periods as the input of each LSTM unit. The validity of the model is proved by experiments of serial-connected LIBs.