Industry-Academia Cooperation

Kozo Keikaku Engineering Inc. established the Social Cooperation Program “Mathematical Engineering for Complex Social Systems in Future” with the Institute of Industrial Science, the University of Tokyo, in February 2016, and has since been collaborating with Professor Kazuyuki Aihara. In this article, we interviewed Professor Aihara of the University of Tokyo, and at the same time we asked Shunya Okuno from KKE, who participated in the joint research, to talk about the endeavor to socially implement the research results.

A New Method of Data-driven Future Prediction

Now it is possible for big data to be used to create mathematical models of objects in computers driven by data and make predictions. The flood forecasting system, which was created from the joint research between KKE and me, also enables highly accurate flood forecasting without creating a conventional physical model. 
 This flood forecasting system is based on a theory that by utilizing a large amount of information that are now able to be measured simultaneously with measuring apparatuses and sensors, information of many variables is transferred to improve the prediction accuracy of a specific target variable. Floods can be predicted by simultaneously measuring data such as water levels at various points in a river and precipitation in various points that affects the river, and using the data “well”. The “using well” part is where algorithm takes part in. By putting the observed data into our algorithm, we can almost automatically predict the specific target variable, such as the water level at a specific point of the river. This has a very wide application range that could be used not only for flood forecasting but also for various other predictions. Furthermore, this technique can handle unprecedented high values, namely extrapolation.

-Established Social Cooperation Program “Mathematical Engineering for Complex Social Systems in Future” with Professor Aihara-

Participation in Social Cooperation Program

 The reason I joined the Social Cooperation Program in 2016, was because of the awareness of the issues concerning the analysis services that I was in charge of. Back then, numerical simulation was constantly advancing in order to solve more complex and large-scale problems. However, as modeling became more and more detailed, the input data on which it was based often contained large observation errors and uncertainties. Under these circumstances, I asked myself “Could I create a new paradigm that could easily and accurately predict from data?” and started my research at Aihara laboratory. 

 

Studying at Aihara Laboratory

Aihara laboratory, where I spent about three years, conducted research at an extremely high level, and everyone was very eager to contribute as much as possible to top journals. Equipped with the common language that is mathematical formulas and English, mathematical experts were challenging creative researches beyond the boundaries of their fields. One time, I was pointed out an error in a complicated mathematical formula on a presentation slide. That was when I realized how researchers see the world differently depending on their proficiency in the common language, and I was astonished.
 One advice I received at the laboratory was to “have an axis that is ‘orthogonal’ to the trend”. While valuing the trends at the time such as deep learning, I dared to pay attention to the dynamical system, which was my “orthogonal axis”. By applying dynamical system theory, the time series data can be analyzed and predicted through a different approach from machine learning. At Aihara laboratory, I researched predictions applying dynamical system, asking questions such as “How should we select and integrate predictions on the spot?”, “How do we predict with a small amount of data?”, or “How do we deal with extrapolation that machine leaning is not good at?”. Based on the results, we managed to develop a flood forecasting system “RiverCast” in 2019 at our Meteorological Disaster Prevention Business Office. 

Social Implementation of Research Results

“RiverCast” is a cloud-based flood forecasting system that makes real-time predictions of water level in river up to 15 hours ahead. In the past, flood forecasting required a lot of time and effort, but by applying the research results from Social Cooperation Program and providing them as a cloud system, we are now able to offer highly accurate flood forecasting easily. Errors in the weather forecast became a huge issue for social implementation, but by modeling errors from past weather forecast data, we made it possible to quantitatively calculate how much uncertainty there is in the predicted water level, and how likely it is to exceed the standard water level.