A dual-feature channel deep network with adaptive variable weight reconstruction for urban water demand prediction

Abstract

Accurate urban water demand forecasting is essential for the rational allocation of water resources and production scheduling, and it is the fundamental guarantee for the safe and efficient operation of water supply systems. However, abrupt changes in exogenous variables can alter user water consumption patterns, introducing uncertainty into urban water demand forecasting. Additionally, fluctuations in residential water demand result in substantial historical noise and mixed variation characteristics within the data, posing challenges to accurate prediction. To address this issue, we propose a frequency improved Legendre memory model with linear-nonlinear dual feature channels, aiming to achieve precise water demand forecasting by thoroughly learning the regularity of urban water usage. In the data instantiation module, we introduce an adaptive gating mechanism to reallocate variable weights, establishing an accurate mapping relationship between exogenous variables and the target variable. Meanwhile, the model's dual-channel structure captures both linear and nonlinear features from historical data, reducing feature extraction complexity while suppressing high-frequency noise. Validated by real urban water demand datasets, this model demonstrates superior performance, achieving accurate water demand forecasting in multiple scenarios. Additionally, the model exhibits strong applicability, consistently achieving the best performance across multiple tasks in public datasets.