Quickly predicting the temperature distribution of a battery pack equipped with sparse temperature sensors is vital in evaluating performance and designing structure. However, limitations of sparse temperature signals, large-scale stack, and complex spatiotemporal characteristics make existing models fail to accurately predict each cell's temperature in the ESS. This paper introduces a spatial-temporal model that quickly predicts the temperature field of the 40-string battery pack with a cell-level computational consumption using the collected sparse signals, where the prior knowledge of battery mechanisms and complex physical modeling are no longer required. The summarized sensor location selection principles ensure the model stays optimal. The predictive performances of the proposed model are discussed in relation to different operating conditions, temperature sensor numbers, and locations. The results show that the sensor-to-cell ratio can be reduced from 1/10 to 1/40 with different training-testing datasets by adhering to the proven principles of optimal sensor position selection. Given the special package structure, the proposed model remains highly effective and accurate in predicting the temperature field of battery packs with one temperature sensor, while the other three benchmarks have a terrible prediction. The proposed engineering-affinity model and ideas are of high practical significance in designing thermal management structures and developing health management strategies for battery systems.