CNN-PatchTST for urban water consumption prediction based on feature selection and multi-scale decomposition

ABSTRACT

With the acceleration of urbanization, urban water consumption prediction has become increasingly critical for refined water resource management. However, existing forecasting methods face three key challenges: insufficient systematic integration of multi-factor influences on water consumption, inadequate exploration of complex temporal structures in water use sequences, and limited capability in capturing local fluctuations in water consumption. To address these gaps, this paper proposes a hybrid forecasting model that integrates Convolutional Neural Networks (CNN) with Patch-based Time Series Transformer (PatchTST), incorporating feature selection and multi-scale decomposition. Specifically, Pearson Correlation Coefficient and Maximal Information Coefficient are combined to screen influential variables, and the water consumption time series is decomposed into trend, seasonal, and residual components. The CNN is embedded into the PatchTST architecture to enhance the extraction of short-term fluctuation features through its local perception capability. Using water consumption data from a residential community in a northern Chinese city, comparative and ablation experiments are conducted. Experimental results demonstrate that the proposed model achieves optimal performance across all metrics, with reductions of 48.5, 61.7, and 56.0% in MSE, MAE, and MAPE, respectively, providing effective technical support for urban water resource management.