Page 17 - 2025年第56卷第7期

（1）：74-77.
［ 26］ 冉笃奎，李敏，武晟，等 . 人工神经网络在径流影响因子滞后性研究中的应用［J］ 计算机工程与应用，
.
2009，45（30）：242-244.
［ 27］ FENG Z，NIU W，TANG Z，et al. Monthly runoff time series prediction by variational mode decomposition and sup⁃
port vector machine based on quantum-behaved particle swarm optimization［J］ Journal of Hydrology，2020，583：
.
124627.

Coupling meteorological information and interpretable deep learning
methods for daily runoff forecasting

1
1
1

LIAO Shengli ，WANG Zaineng ，CHENG Chuntian ，WU Huijun 2

（1. Institute of Hydropower & Hydroinformatics，Dalian University of Technology，Dalian 116024，China；

2. Electric Power Dispatching Control Center，China Southern Power Grid Company Limited，Guangzhou 510663，China）

Abstract：Accurate runoff forecast information is an important basis for flood control and hydropower scheduling.

However，factors such as the selection of input features，determination of model structure，and interpretability of
forecasting results severely limit the application of data-driven models in runoff forecasting. This study utilizes multi-
source data，including observed runoff，rainfall，and ERA5 reanalysis data enriched with comprehensive meteorologi⁃

cal information，as inputs. The framework combines the SHapley Additive exPlanations （SHAP）method with a Con⁃
volutional Neural Network-Bidirectional Long Short-Term Memory （CNN-BiLSTM）model to establish an interpre⁃

table runoff forecasting framework optimized through a post-hoc feature selection mechanism. Firstly，the multi-

source data enriches the framework with diverse input information，with CNN and BiLSTM respectively capturing spa⁃

tial correlations and temporal patterns in the data to enhance forecasting accuracy. Secondly，following model train⁃
ing，SHAP calculates the contribution of each input feature. By comparing model performance under different input

conditions，optimal input features and model structures are determined. Finally，by quantifying and visualizing the

contribution of input features，the framework provides insights into the forecasting mechanism and results. The pro⁃

posed method is applied to interval inflow forecasting at the Tian Sheng Qiao Hydropower Station and compared with
models using input features selected by traditional methods such as Partial Autocorrelation Function （PACF），Cross-

.
Correlation Function （CCF），and Random Forest （RF） The results demonstrate that the SHAP-based approach not

only optimizes feature selection compared to traditional methods，but also significantly improves the challenges faced

by data-driven models in runoff forecasting，including difficulties in input feature selection，inefficiencies，and inter⁃

preting forecasting results.

Keywords：runoff forecasting；SHAP；CNN-BiLSTM；ERA5；input features；interpretability

（责任编辑：耿庆斋）

— 843 —