Previsão de demanda de curto prazo em sistemas de abastecimento de água empregando modelos estatísticos, de aprendizagem de máquina e de varredura

Abstract

In the operation of Water Supply Systems (SAA), one of the critical factors of supply is maintaining the balance between water supply and demand to users. The maintenance of this balance is made through operational actions and the forecast of short term demand is a major factor for this management. Over the years, numerous methods have been developed and tested. Linear models have been widely used because they are easy to develop and implement, and simple to understand and interpret. However, water demand data have different degrees of nonlinearity, which cannot be adequately addressed by linear models. The literature review highlights that the most successful models are based on soft computing approaches such as neural networks, fuzzy systems, evolutionary computing, support vector machines, and hybrid models. In this paper we investigate the feasibility of using the Dynamic Time Scan Forecasting (DTSF) time scan method to predict short-term (hourly) demand in water supply systems by comparing efficiency and computational cost (CC), with several known univariate alternatives. The statistical and neuronal methods employed were: Box-Jenkins (SARIMA), Exponential Smoothing (ETS), Hybrid model using Seasonal and Trend decomposition using Loess filter (STL) and ETS (STL-ETS), Trigonometric Box-Cox transformation model, ARMA errors, Trend and Seasonal components (TBATS), in addition to hybrid neuronal models with bootstrap nonlinear autoregressive neural network (BNNAR), Extreme Learning Machine (ELM) and Naive Bayes (NB). The data used in this study refer to a supply zone (ZA) located in the south central region of a capital of south eastern Brazil and serves a population of approximately 230,000 people, and can be compared with a medium-sized city. In order to evaluate the methods, the following metrics were employed: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Model Fitting (MF), in addition to WC. The results show that in both summer and fall there was no model that outperformed the others in terms of forecast accuracy. However, for the months from June to November (winter and spring), the ELM method showed the best results. The two neuronal methods showed the best accuracy, however, the greatest CC when employing cross-validation of sliding origin with recalibration. Regarding the CC statistic, the DTSF algorithm was extremely fast compared to the other methods studied for all months of the year, using a fraction of the time. The results obtained showed that the proposed method provides similar or improved forecast values compared to the soft computing and statistical methods employed, but using a fraction of the computation time. The great advantage of this method, as it is a data-driven method, is that the more data you use, the better your generalizability will be, and in an environment known as the big data era, it will take advantage of others methods.