In this study, we will construct a control system that comprehensively evaluates the impact on runoff floodwater resources and society, targeting scenarios that effectively suppress heavy rainfall by implementing multiple engineering methods, at multiple points in time and stages (Figure 1). Additionally, the control system would be designed to intervene at multiple stages to course-correct when an unexpected deviation occurs. Furthermore, our goal is to implement multiple types of interventions at various stages to increase the regulatory effect. Specifically, we aim to develop a decision-making support system that can derive optimal solutions by combining multiple control methods in real-time by (1) simplifying the time evolution model for heavy rain events and constructing an ensemble prediction method, (2) constructing monitoring methods for regulation, (3) setting appropriate objective functions based on the output of ELSI/RRI research, and (4) optimizing algorithms.

① Demonstration of a Real-Time decision support system

The objective of R&D Item 2 is to develop a system capable of deriving optimal solutions for operations in real-time. This system integrates knowledge from Item 1 on various operational methods and from Item 3 on the evaluation of impacts on floods, water resources, and human societies.
To create a demonstration of the real-time decision support system, we produced a demonstration video based on a feasible heavy-rainfall control scenario. Specifically, in collaboration with other research and development tasks within the project, we seamlessly reproduced real-time situations in a three-dimensional virtual environment based on the results of current numerical simulations, including decision-making for implementing heavy-rainfall control, selection of seeding locations, deployment of seeding aircraft, seeding of rain clouds, changes in rain clouds, river water levels, inundation depth, and the corresponding estimated damage.

② Verification of the effectiveness of the ensemble mean adjoint method

In research on actuator location optimization, it has become evident that conventional adjoint models are not very effective for nonlinear optimization in meteorological models. This is suggested to be because the chaotic nature of the atmosphere makes the optimization prone to becoming trapped in local minima. To address this issue, we investigated the use of an ensemble mean adjoint method, in which adjoint calculations are performed for multiple cases with added perturbations, and the resulting sensitivities are averaged to obtain a more robust sensitivity estimate.
For a WRF simulation of the Western Japan Heavy Rain Event, impulse inputs were applied and analyzed based on sensitivities derived from the conventional adjoint equations in order to reduce the accumulated precipitation within the red-framed area shown in Figure 3. The results show changes in accumulated rainfall up to six hours later when surface humidity was modified according to the sensitivities while keeping the input intensity constant. Compared with the case without ensemble averaging, the introduction of appropriate random perturbations enabled a significant reduction in accumulated rainfall.