Evaluation of a Multivariate Spatiotemporal Processing Model Incorporating Local Inductive Bias in Inter‐Dam Transfers

To improve the accuracy of dam inflow prediction under limited data conditions, we propose inter‐dam transfer learning using SD‐MTSM (Spatially Distributed Multivariate Time Series Model). This method adds a spatial processing layer with local inductive bias to the multivariate time series model TSMixer to account for differences in spatial domains between dams. Experimental results show that the proposed method achieves high accuracy compared to the base model TSMixer in scratch learning with limited data. Furthermore, in fine‐tuning, high accuracy is achieved when the target dam's catchment area is small and the source's catchment area is large.