Advanced flood risk prediction using AI technology: CTI Engineering Co., Ltd. has developed a real-time automatic learning function for flood prediction system models.

To efficiently improve the precision of risk information related to small and medium-sized rivers for which there is not enough observational data, we have developed an automatic learning function that utilizes new flood-related observational data obtained while operating a flood prediction system. This technology will enable the prompt precision improvement of risk information related to small and medium-sized rivers as well as river basins, including upper river basins.

1. Background
Because improving the precision of flood prediction system models requires advanced technical judgment that considers a huge amount of case verification work and the specific characteristics of each river, this has been limited to large rivers and the midstream and downstream portions of rivers that affect urban areas until now, which means that it has been difficult to improve the precision related to small and medium-sized rivers and upper river basins.

However, due to abnormal flooding in recent years, even small and medium-sized rivers, etc. have frequently caused flooding damage, so it is becoming increasingly important to provide highly precise risk information related to rivers for which there is little observational data.

　

2. Overview and effects of the developed technology
(1) Overview of the developed technology
① Automatic flood-prediction-model learning

• We have developed an automatic learning function that utilizes river-water-level observational data obtained using IoT devices, etc. (Figure 1)
• This automatic learning function utilizes observed river-water-level data in combination with AI technology to enable the prompt setting of optimal flood-predication-model parameters after floods end.　

　

Figure 1. Effect of introducing our automatic learning function

　

② Incorporation of our automatic learning function into a flood prediction model linked to RisKma

• By adding a flood prediction model that incorporates our newly developed automatic learning function to RisKma--the water-related disaster risk mapping system we provide to municipalities--as a new function, it can be used alongside real-time water-related disaster monitoring functions that include predicted rainfall, river water levels, and flood risks.

　

(2) Effects of the developed technology (increased flood-prediction precision and efficiency thanks to the automatic learning function)

• Because parameters are automatically optimized by using new flood observational data incorporated via IoT devices, etc., the flood-prediction precision can be improved. (Figure 2)
• The automatic learning function can be used to speed up the verification and improvement process related to increasing flood-prediction-model precision.
• In terms of observatories located at upper river basins and tributaries for which there is little past observational data and crisis-management water level meters--which have been increasingly installed in recent years--as well, a flood prediction model can be started up as a predictive system by using general values and other provisional parameters, and then the precision can be automatically improved after the occurrence of floods.

　

Figure 2. Improving reproducibility via additional learning

　

3. Future prospects
We will strive to promote the introduction of our technology in areas that lack flood-disaster risk information, resolve issues that arise during operation, and thereby contribute to water-related disaster prevention throughout society by providing flood-disaster risk information.