Integrated Data Preprocessing Strategies for Improved Water Demand Prediction: A Case Study with the Informer Model

Predicting water demand poses a challenge for water utility companies, primarily due to data loss and anomalies that occur during the data collection process. However, most research on water quantity prediction tends to focus more on constructing models rather than on data processing. The main objective of this study is to tackle this issue by preprocessing the dataset to improve the predictive capability of the model. This is achieved by maximizing the resolution of anomalies, missing values, and noise. The methods employed in this study involve various data preprocessing techniques applied to time series data, including outlier handling, missing value imputation, and noise smoothing. For missing value imputation, a method combining Random Forest and K-Nearest Neighbor (RF-KNN) is proposed. Additionally, the preprocessed data is fed into the Informer model. Through optimal preprocessing, there is a noticeable improvement in the evaluation metrics. The optimal preprocessing combinations have resulted in an 85.13% improvement in Mean Squared Error (MSE) and an 83.86% improvement in different models. However, the Mean Absolute Error (MAE) has increased by 42.17% and 37.97% respectively. Comparative analysis and discussions further demonstrate that the ensemble preprocessing methods of Interquartile Range (IQR), RF-KNN, and Fourier Transform (FT) significantly enhance the predictive accuracy and reliability of the model when applied to water datasets.