Evaluation of Various Machine Learning Methods to Predict Istanbul’s Freshwater Consumption

Abstract

Planning, organizing, and managing water resources is crucial for urban areas and metropolitans. Istanbul is one of the largest megacities, with a population of over 15 million. The large volume of water demand and increasing scarcity of clean water resources make long-term planning necessary for this city, as sustained water supply requires large-scale investment projects. Successful investment plans require accurate projections and forecasting for freshwater demand. This study considers different machine learning methods for freshwater demand forecasting for Istanbul. Using monthly consumption data provided by the municipality since 2009, we compare forecasting accuracies of ARIMA, Holt-Winters, Artificial Neural Networks, Recursive Neural Networks, Long-Short Term Memory, and Simple Recurrent Neural Network models. We find that the monthly freshwater demand of Istanbul is best predicted by Multi-Layer Perceptron and Seasonal ARIMA. From the predictive modeling perspective, this result is another indication of the combined usage of conventional forecasting models and novel machine learning techniques to achieve the highest forecasting accuracy.