Optimization of Support Vector Regression Models for Stormwater Prediction

In this paper we propose a solution to a real-world time series regression problem: the prediction of fill levels of stormwater tanks. Our regression model is based on Support Vector Regression (SVR), but can easily be replaced with other data mining methods. The main intention of the work is to overcome frequently occuring problems in data mining by automatically tuning both preprocessing and hyperparameters. We highly believe that many models can be improved by a systematic preprocessing and hyperparameter tuning. The optimization of our model is presented in a step-by-step manner which can easily be adapted to other time series problems. We point out possible issues of parameter tuning, e.g., we analyze our tuned models with respect to overfitting and oversearching (which are effects that might lead to a reduced model generalizability) and present methods to circumvent such issues.