• time series
• random forest
• machine learning
• urban flood forecasting

Predicting the maximum possible level of water in junctions found in the urban drainage model can be a great value in forecasting urban flooding. Monitoring and predicting natural environments can be a challenging task by considering the variable behavior of the domain as well as the hostile condition. Since the degree and scale of the flood hazards and damages increases over time with the changing climate, the issues need to be addressed without further delay. To tackle the challenge one should consider involving knowledge from different disciplines like hydrology, meteorology, statistics and data science etc. to mention a few. In this paper we are going to propose and discuss a methodology to predict level of water in a junction found in urban drainage model in particular in the ancient town of Pisa, Italy. Specifically the approach is to develop and train a machine learning model that can forecast the level of water in the junction under study with different lead time. To do this we used historical time series data of rain rate with its corresponding Inflows on the junction under study as a learning set and applying different machine learning algorithms based on heuristic criteria as a methodology.
However the main ingredient for this task which is the historical dataset representing the level of water in junctions with the corresponding precipitation is absent. Therefor there should be a mechanism to prepare this dataset. To prepare the historical dataset we have used the SWMM5 simulator and generate the dataset synthetically. By applying the precipitation records collected with fifteen minute interval for around three years and seven months together with the hydrological model of the area under study which is provided by Aqua s.r.l to the SWMM5 simulator we are able to generate the historical dataset needed in order to build and train the predictive model.
By using the result of the simulation which produces a time series data that show the level of water in each junction at the specified time series we are able to develop and train the predictive model. To build the model we have used the Java API of Weka. We implement the predictive model using three different algorithms which are Linear Regression, Decision tree and Random forest. Finally we have done experiments to assess the performance of the model implemented using these algorithms.