Challenge
As a municipal utility, the customer is responsible, among other things, for the entire water supply of a city in Germany. Accurate forecasts of water requirements are essential for maintenance measures and water redistribution in the event of pipe bursts, for example. For example, water redistribution measures in the event of pipe bursts require a precise assessment of water requirements at any time of day. Although demands follow strong daily and weekly seasonality (given by working days and demands in industry), the influence of weather and other factors cannot be neglected. At the start of the project, the customer creates the forecasts manually by hand and only checks the influences of external factors on a random basis. In addition to the challenges in forecasting, there were also difficulties with the quality of the data. For example, measurement values were historically collected manually, which led, for example, to missing values when individual employees were ill. The aim of the project was therefore to predict the hourly water requirements for the next seven days and the daily water requirements for the next 30 days as precisely as possible.
Approach
In addition to historical data on water consumption over the last five years, various external influencing factors were collected to predict water demand volumes, such as weather data (temperature, rainfall), school and semester vacations, other calendar information such as seasons and public holidays, and special events (such as soccer matches). In particular, the weather data was used to derive further variables: For example, the days since the last rainfall, the amount of precipitation in the last week or the days until the next precipitation can be used. Based on water demand history and other external factors, machine learning forecast models were trained to calculate future water requirements. Of the influencing factors evaluated, weather data and calendar information ultimately had the biggest influence on forecasts.
Outcomes
With the help of machine learning models, hourly and daily water requirements could be predicted with a very high level of accuracy. That is, with an accuracy of 95.0% for hourly forecasts and an accuracy of 94.6% for daily forecasts. As a result, the predictions were 32% and 43% more accurate than an individually constructed benchmark that is not based on machine learning algorithms. As a result of more accurate demand forecasts, the municipal utility benefits from increased planning security for maintenance measures or even water redistribution, for example.