Unbiased Time Series Forecasting for Sustainable Water Management

In recent years, many municipalities and communities have begun to utilize sensors in their water management procedures, while many others aim to do so in the near future, in order to improve the provided services to their citizens, based on the data generated by the abovementioned sensors. Those data, which are time series data, can either be used for monitoring the system or even for training Machine Learning (ML) models for providing predictions based on the said data. The latter is also called time series forecasting. However, the aforementioned data are not always of high quality due to the nature of the source from which they originate. More specifically, a sensor may fail at any given time and start generating erroneous data, thus affecting the accuracy of the corresponding trained ML algorithm that, in return, will provide biased predictions. Moreover, the methods used for the preprocessing and the analysis of the time series data, as well as the ML algorithms themselves that are being utilized, should be the appropriate ones, since they can lead to biased predictions. As a result, the necessary steps should be followed when utilizing this kind of data for training ML models for time series forecasting, alongside the whole data lifecycle. However, addressing the bias issue in its origin means changing the sensors with higher quality ones, which is not always affordable or may not be even necessary, since there are approaches that are able to mitigate bias in the later stage of the data lifecycle. To this end, this paper proposes an approach for performing unbiased times series forecasting on time series data coming from sensors of water management facilities. The proposed approach focuses on the steps coming after the data generation and collection, namely data preprocessing, data analysis, model training and model evaluation.