Reservoir computing presents high performance in processing time-series data with a low computational cost. Among recent proposals, so-called multi-reservoir computing (MRC), which uses multiple reservoirs, has been attracting attention. For example, a series of reservoirs are connected in a hierarchical manner to process signals, or reservoirs having different parameters are connected in parallel to output various signals to be used for generating final outputs. These proposals work effectively by expanding the capabilities of a single reservoir. Meanwhile, artificial intelligence (AI) broadly has made progress in emphasizing data drive. There, the properties in the data often prescribe the structure of a processing system, or a system exhibits higher functionality when its structure is aligned with the data properties. This paper proposes a multi-reservoir system for spatiotemporal forecasting of human flows, or local population evolution, where the coupling topology of multiple reservoirs is consistent with the geography in real space. We name this the geo-reservoir computing (GeoRC) system. It can effectively utilize the structure of the real space from which the data is acquired. Experimental results show that the GeoRC system can forecast the spatiotemporal changes of the local population with higher accuracy than conventional methods. This paper not only proposes a method to achieve highly accurate spatiotemporal forecasts, but also suggests that the structure contained in the real world, where data is produced, is closely related to the structure possessed by neural networks that handle spatiotemporal data, which is of great significance in future data-driven neuro information processing.