Water Supply Network Flow Rate Prediction for Short-Duration by STL-LSTM

Accurate prediction of flow in water distribution networks is essential for improving the operational efficiency of urban water supply systems and optimizing resource allocation. To address this need, this paper presents a short-duration flow prediction model for user nodes within a water distribution network. The proposed approach integrates seasonal-trend decomposition (STL) of flow sequences with a long short-duration memory (LSTM) network enhanced by multi-feature inputs. The modeling process begins with linear interpolation to preprocess the flow sequence, ensuring data continuity. The STL method is then employed to decompose the sequence into trend, seasonal, and residual components. Each component is modeled separately using dedicated LSTM networks to capture its unique temporal characteristics. The predicted components are subsequently aggregated to reconstruct the final short-term flow forecast. To evaluate model performance, residual analysis is conducted, confirming the robustness of the proposed approach. Furthermore, comparative experiments are performed against benchmark models including CNN, LSTM, and STL-CNN. Results show that the STL-LSTM model achieves a reduction in forecasting error of over 3% on half-hourly data, and it is verified that STL is not suitable for combining with CNN.