Predicting the Overflowing of Urban Personholes Based on Machine Learning Techniques

Abstract

Urban stormwater drainage systems, which include many personholes to collect and discharge precipitation within a city, are extensively constructed to prevent streets and buildings from flooding. This research intends to build a machine learning model to predict whether a personhole will overflow soon, which is crucial to alleviate the damage caused by floods. To address the challenges posed by many diverse personholes, we proposed segmenting the personholes into several groups and have designed two methods employing different personhole features. The first, the geography-based method, uses the geographical locations of the personholes for the grouping. The second, the hydrology-based method, uses the characteristics that are directly related to the overflowing situation, such as the depth of the personhole, and the average and the maximum water level of the personholes. We also investigated several machine learning techniques, such as the multilayer perceptron (MLP) model and a fine-tuning architecture. The study area was located in the new Taipei city and the experimental results have shown the impressive predictive ability of the proposed approaches. Particularly, by applying the hydrology-based grouping method, and using a hybrid model combining the machine learning model prediction results with heuristic rules, we can obtain the best prediction result, and the accuracy is over 99%. We have also noticed the influence of the activation function used in the neural network and the number of frozen layers in the fine-tuning architecture. Particularly, using the tanh function with one frozen layer is good in some cases. However, since it is not general enough, we suggest the readers perform empirical studies before choosing the best setting in their own environment.