The importance of efficient water resource supply has been acknowledged, and it is essential to predict short-term water consumption in the future. Recently, it has become possible to obtain data on water consumption at the household level through smart water meters. The pattern of these data is nonlinear due to various factors related to human activities, such as holidays and weather. However, it is difficult to accurately predict household water consumption with a nonlinear pattern with the autoregressive integrated moving average (ARIMA) model, a traditional time series prediction model. Thus, this study used a deep learning-based long short-term memory (LSTM) approach to develop a water consumption prediction model for each customer. The proposed model considers several variables to learn nonlinear water consumption patterns. We developed an ARIMA model and an LSTM model in the training dataset for customers with four different water-use types (detached houses, apartment, restaurant, and elementary school). The performances of the two models were evaluated using a test dataset that was not used for model learning. The LSTM model outperformed the ARIMA model in all households (correlation coefficient: mean 89% and root mean square error: mean 5.60 m³). Therefore, it is expected that the proposed model can predict customer-specific water consumption at the household level depending on the type of use.

Owing to global climate change, the frequency of disasters has increased twelve-fold, with a corresponding approximately seventeen-fold increase in economic damages over the past six decades. Notably, severe flood damage has been occurring in Asia along the Pacific coast due to extreme weather events, including torrential rains and typhoons, which have been becoming increasingly frequent and prolonging the rainy season. In the Eastern Visayas region, the management and monitoring facilities for flood observation data, as well as the forecasting and warning systems suitable for the local area, are insufficient. The warning system introduced through overseas grants is limited in operation in some areas of the city. Furthermore, although an organization has jurisdiction over flood forecasting and warning, the system’s operation is not systematic and is limited. Additionally, there is a shortage of technical manpower. In this study, we utilized deep learning models to forecast flood water levels in the CarayCaray Basin on Biliran Island, located in Eastern Visayas, the Philippines. Additionally, a flood risk classification was applied to evaluate the degree of risk associated with the predicted water levels. The predicted water levels for each model were compared with the observed water level data. The evaluation of the predictive performance of each model resulted in an NRMSE value of 9.48. Moreover, the accuracy of the DNN model was found to be the best among the flood water level prediction models. To implement the flood risk classification, we utilized extreme gradient boosting, random forest, and decision tree models. The application of these models resulted in an F1-score of 0.92 for the extreme gradient boost model, which exhibited the highest accuracy. With an increasing need for disaster (flood) management, AI-based predictive models are anticipated to reduce the damage caused by natural disasters and enhance disaster mitigation systems. Real-time collection of rainfall and water level data enables continuous learning. Furthermore, if a clear flood warning based on learned flood level patterns is issued, preemptive measures can be taken before intense flood damage occurs.

Vietnam faces on significant human and property losses from floods almost every year. Therefore, the aim of this study is to provide timely and highly accurate flood prediction information using a developed hybrid model by combination of the rainfall-runoff model and AI-based model. We used Tank model as a rainfall-runoff model for peak flood discharge and parameter calibration was conducted by GA (genetic algorithm) and PS (pattern search). The sum of squared residuals (SSR) and weighted sum of squared residuals (WSSR) as objective functions were used for peak flood discharge evaluation. The simulated flood discharge was converted as flood water level by rating curve and Monte Carlo simulation was applied to estimate the confidence bounds for 95% confidence level. These estimates were then utilized as input data for the AI-based models to identify the strengths and weaknesses of each model and develop an optimal flood water level prediction model. Two AI-based models, deep neural network (DNN) and long short-term memory (LSTM), were used for flood water level prediction. The LSTM model demonstrated the best performance with a correlation coefficient (CC) of 0.98, normalized root mean square error (NRMSE) of 0.04, and Nash-Sutcliffe efficiency (NSE) of 0.98.

To improve flood forecasting accuracy, this study proposes a hybrid model that combines a physically based rainfall-runoff model with AI-based residual prediction results. The storage function model was used to simulate runoff, while AI models (random forests, support vector regression, long short term memory, and gated recurrent unit) were used to predict the residuals. The hybrid model calculates corrected runoff by combining simulated runoff from the storage function model with AI-predicted residuals. Compared with the standalone storage function model and AI models for runoff prediction, the hybrid model demonstrated effectiveness in predicting both peak discharge and the timing of peak discharge. The proposed hybrid model can improve flood forecasting reliability and offers a valuable tool for early warning and disaster management.

Korea’s early warning system for flood disaster management is starting preparations for flood response by the heavy rain advisory (HRA) of the Korea Meteorological Administration (KMA). However, the HRA criterion has a critical limitation in that it considers only consistent rainfall patterns (e.g. 60 mm/3 hrs or 110 mm/12 hrs) without considering the characteristics of heavy rain damage in the region. To address this problem, the present study proposes a heavy rain damage prediction technique based on optimization and ensemble method. To predict damage as accurately as possible, fifteen rainfall variables according to the duration and magnitude are considered. The dataset is divided into a training dataset (70%) and a test dataset (30%) by random extraction. An optimal threshold that the damage can be occurred is derived for each region via optimization method of the training dataset. The area under the receiver operating characteristic (AUROC) curve, F1 score, and F2 score are each considered as objective function, and the F2 score was selected because it is more effective in terms of disaster management. In addition, the method is designed to predict damage probabilistically by applying the ensemble concept. This novel method is defined as a heavy rain damage prediction technique (HDPT). Finally, the HDPT is evaluated using the test dataset and by comparison with the results from the HRA data, and the F2 score of the HDPT is shown to be about 10% higher than that of the HRA. Thus, the proposed methodology is expected to be more effective than the current HRA method for the early warning system and for disaster management.