Improving Daily Runoff Prediction Using a Novel Two‐Step Post‐Processing Method of Frequency Distribution Curve Correction

Abstract

Offsets and distributional discrepancies between runoff simulations and observations were two important error sources which remarkably degraded runoff prediction performance of hydrological models. However, traditional single post‐processing correction methods were difficult to reduce these two error sources simultaneously due to their specific optimization objectives. Our study proposed a two‐step post‐processing method that integrated multi‐statistical (i.e., Linear scaling: LS, Quantile mapping: QM, and Bernoulli‐Gamma‐Gaussian: BGG) and machine‐learning (i.e., Long Short‐Term Memory networks: LSTM) methods for a watershed hydrological model. The efficiency coefficient (NSE) and Jensen‐Shannon divergence (JS) were adopted to evaluate the runoff prediction performance with different frequencies (i.e., high: ≥25th percentile of flow duration curve, medium: ≥75th and <25th, and low: <75th) and distributional discrepancy, respectively. Eight watersheds were selected as our study area to test the robustness of our proposed method by comparing with those of single post‐processing methods (i.e., LS, QM, BGG and LSTM). Results showed that our proposed method achieved considerable improvements for all the flow predictions, particularly for QM‐LSTM. The best improvements were for the medium flow, followed by low, entire and high flows, whose NSE values increased by 31.1% ∼ 97.2%, 32.3% ∼ 97.3%, 15.8% ∼ 93.4% and 12.7% ∼ 88.7% with a mean of 67.5%, 65.2%, 51.5% and 46.0%, respectively. The JS values of predicted flow duration curve decreased by 12.0% ∼ 97.2% with a median of 74.3%. Compared with the optimal single method (i.e., LSTM), the mean NSE increased by 6.4%, 6.5%, 26.4% and 43.2% for the entire, high, medium and low flows, respectively, and the mean JS decreased by 20.7%.