Exploring a similarity search-based data-driven framework for multi-step-ahead flood forecasting

Due to a small proportion of observations, reliable and accurate flood forecasts for large floods present a fundamental challenge to artificial neural network models, especially when the forecast horizons exceed the flood concentration time of a river basin. This study proposed for the first time a Similarity search-based data-driven framework, and takes the advanced Temporal Convolutional Network based Encoder-Decoder model (S-TCNED) as an example for multi-step-ahead flood forecasting. A total of 5232 hourly hydrological data were divided into two datasets for model training and testing. The input sequence of the model included hourly flood flows of a hydrological station and rainfall data (traced back to the previous 32 h) of 15 gauge stations, and the output sequence stepped into 1- up to 16-hour-ahead flood forecasts. A conventional TCNED model was also built for comparison purposes. The results demonstrated that both TCNED and S-TCNED could make suitable multi-step-ahead flood forecasts, while the proposed S-TCNED model not only could effectively mimic the long-term rainfall-runoff relationship but also could provide more reliable and accurate forecasts of large floods than the TCNED model even in extreme weather conditions. There is a significant positive correlation between the mean sample label density improvement and the mean Nash-Sutcliffe Efficiency (NSE) improvement of the S-TCNED over the TCNED at the long forecast horizons (13 h up to 16 h). Based on the analysis of the sample label density, it is found that the similarity search largely improves the model performance by enabling the S-TCNED model to learn the development process of similar historical floods in a targeted manner. We conclude that the proposed S-TCNED model that converts and associates the previous rainfall-runoff sequence with the forecasting runoff sequence under a similar scenario can enhance the reliability and accuracy of flood forecasts while extending the length of forecast horizons.