Flood simulation and forecasting in various types of watersheds is a hot issue in hydrology. Conceptual hydrological models have been widely applied to flood forecasting for decades. With the development of economy, modern China faces with severe flood disasters in all types of watersheds include humid, semi-humid semi-arid and arid watersheds. However, conceptual model-based flood forecasting in semi-humid semi-arid and arid regions is still challenging. To investigate the applicability of conceptual hydrological models for flood forecasting in the above mentioned regions, three typical conceptual models, include Xinanjiang (XAJ), mix runoff generation (MIX) and northern Shannxi (NS), are applied to 3 humid, 3 semi-humid semi-arid, and 3 arid watersheds. The rainfall-runoff data of the 9 watersheds are analyzed based on statistical analysis and information theory, and the model performances are compared and analyzed based on boxplots and scatter plots. It is observed the complexity of drier watershed data is higher than that of the wetter watersheds. This indicates the flood forecasting is harder in drier watersheds. Simulation results indicate all models perform satisfactorily in humid watersheds and only NS model is applicable in arid watersheds. Model with consideration of saturation excess runoff generation (XAJ and MIX) perform better than the infiltration excess-based NS model in semi-humid semi-arid watersheds. It is concluded more accurate mix runoff generation theory, more stable and efficient numerical solution of infiltration equation and rainfall data with higher spatial-temporal resolution are main obstacles for conceptual model-based flood simulation and forecasting.

The coupled routing and excess storage (CREST) distributed hydrological model has been applied regionally and globally for years. With the development of remote sensing, requirements for data assimilation and integration have become new challenges for the CREST model. In this paper, an improved CREST model version 3.0 (Tsinghua University and China Institute of Water Resources and Hydropower Research, Beijing, China) is proposed to enable the use of remotely-sensed data and to further improve model performance. Version 3.0 model’s runoff generation, soil moisture, and evapotranspiration based on three soil layers to make the CREST model friendly to remote sensing products such as soil moisture. A free water reservoir-based module which separates three runoff components and a four mechanism-based cell-to-cell routing module are also developed. Traditional CREST and CREST 3.0 are applied in the Ganjiang River basin, China to compare their simulation capability and applicability. Research results indicate that CREST 3.0 outperforms the traditional model and has good application prospects in data assimilation, flood forecasting, and water resources planning and management applications.

Climate change and human activities have a great impact on the environment and have challenged the assumption of the stability of the hydrological time series and the consistency of the observed data. In order to investigate the applicability of machine learning (ML)-based rainfall–runoff (RR) simulation methods under a changing environment scenario, several ML-based RR simulation models implemented in novel continuous and non-real-time correction manners were constructed. The proposed models incorporated categorical boosting (CatBoost), a multi-hidden-layer BP neural network (MBP), and a long short-term memory neural network (LSTM) as the input–output simulators. This study focused on the Dongwan catchment of the Yiluo River Basin to carry out daily RR simulations for the purpose of verifying the model’s applicability. Model performances were evaluated based on statistical indicators such as the deterministic coefficient, peak flow error, and runoff depth error. The research findings indicated that (1) ML-based RR simulation by using a consistency-disrupted dataset exhibited significant bias. During the validation phase for the three models, the R² index decreased to around 0.6, and the peak flow error increased to over 20%. (2) Identifying data consistency transition points through data analysis and conducting staged RR simulations before and after the transition point can improve simulation accuracy. The R² values for all three models during both the baseline and change periods were above 0.85, with peak flow and runoff depth errors of less than 20%. Among them, the CatBoost model demonstrated superior phased simulation accuracy and smoother simulation processes and closely matched the measured runoff processes across high, medium, and low water levels, with daily runoff simulation results surpassing those of the BP neural network and LSTM models. (3) When simulating the entire dataset without staged treatment, it is impossible to achieve good simulation results by adopting uniform extraction of the training samples. Under this scenario, the MBP exhibited the strongest generalization capability, highest prediction accuracy, better algorithm stability, and superior simulation accuracy compared to the CatBoost and LSTM simulators. This study offers new ideas and methods for enhancing the runoff simulation capabilities of machine learning models in changing environments.

For the purpose of improving the scientific nature, reliability, and accuracy of flood forecasting, it is an effective and practical way to construct a flood forecasting scheme and carry out real-time forecasting with consideration of different rain patterns. The technique for rain pattern classification is of great significance in the above-mentioned technical roadmap. With the rapid development of artificial intelligence technologies such as machine learning, it is possible and necessary to apply these new methods to assist rain classification applications. In this research, multiple machine learning methods were adopted to study the time-history distribution characteristics and conduct rain pattern classification from observed rainfall time series data. Firstly, the hourly rainfall data between 2003 and 2021 of 37 rain gauge stations in the Pi River Basin were collected to classify rain patterns based on the universally acknowledged dynamic time warping (DTW) algorithm, and the classifications were treated as the benchmark result. After that, four other machine learning methods, including the Decision Tree (DT), Long- and Short-Term Memory (LSTM) neural network, Light Gradient Boosting Machine (LightGBM), and Support Vector Machine (SVM), were specifically selected to establish classification models and the model performances were compared. By adjusting the sampling size, the influence of different sizes on the classification was analyzed. Intercomparison results indicated that LightGBM achieved the highest accuracy and the fastest training speed, the accuracy and F₁ score were 98.95% and 98.58%, respectively, and the loss function and accuracy converged quickly after only 20 iterations. LSTM and SVM have satisfactory accuracy but relatively low training efficiency, and DT has fast classification speed but relatively low accuracy. With the increase in the sampling size, classification results became stable and more accurate. Besides the higher accuracy, the training efficiency of the four methods was also improved.

Runoff forecasting is crucial for water resource management and flood safety and remains a central research topic in hydrology. Recent advancements in machine learning provide novel approaches for predicting runoff. This study employs the Competitive Adaptive Reweighted Sampling (CARS) algorithm to integrate various machine learning models into a data-driven rainfall–runoff simulation model. We compare the forecasting performance of different machine learning models to improve rainfall–runoff prediction accuracy. This study uses data from the Maduwang hydrological station in the Bahe river basin, which contain 12 measured flood events from 2000 to 2010. Historical runoff and areal mean rainfall serve as model inputs, while flood data at different lead times are used as model outputs. Among the 12 flood events, 9 are used as the training set, 2 as the validation set, and 1 as the testing set. The results indicate that the CARS-based machine learning model effectively forecasts floods in the Bahe River basin. Under the prediction period of 1 to 6 h, the model achieves high forecasting accuracy, with the average NSE ranging from 0.7509 to 0.9671 and the average R² ranging from 0.8397 to 0.9413, though the accuracy declines to some extent as the lead time increases. The model accurately predicts peak flow and performs well in forecasting high flow and recession flows, though peak flows are somewhat underestimated for longer lead times. Compared to other machine learning models, the SVR model has the highest average RMSE of 0.942 for a 1–6 h prediction period. It exhibits the smallest deviation among low-, medium-, and high-flow curves, with the lowest NRMSE values across training, validation, and test sets, demonstrating better simulation performance and generalization capability. Therefore, the machine learning model based on CARS feature selection can serve as an effective method for flood forecasting. The related findings provide a new forecasting method and scientific decision-making basis for basin flood safety.

In the field of hydrological model parameter uncertainty analysis, sampling methods such as Differential Evolution based on Monte Carlo Markov Chain (DE-MC) and Shuffled Complex Evolution Metropolis (SCEM-UA) algorithms have been widely applied. However, there are two drawbacks which may introduce bad effects into the uncertainty analysis. The first disadvantage is that few optimization algorithms consider the physical meaning and reasonable range of the model parameters. The traditional sampling algorithms may generate non-physical parameter values and poorly simulated hydrographs when carrying out the uncertainty analysis. The second disadvantage is that the widely used sampling algorithms commonly involve only a single objective. Such sampling procedures implicitly introduce too strong an “exploitation” property into the sampling process, consequently destroying the diversity property of the sampled population, i.e., the “exploration” property is bad. Here, “exploitation” refers to using good already-existing solutions and making refinements to them, so that their fitness will improve further; meanwhile, “exploration” denotes that the algorithm searches for new solutions in new regions. With the aim of improving the performance of uncertainty analysis algorithms, in this research, a constrained multi-objective intelligent optimization algorithm is proposed that preserves the physical meaning of the model parameter using the penalty function method and maintains the population diversity using a Non-dominated Sorted Genetic Algorithm-II (NSGA-II) multi-objective optimization procedure. The representativeness of the parameter population is estimated on the basis of the mean and standard deviation of the Nash–Sutcliffe coefficient, and the diversity is evaluated on the basis of the mean Euclidean distance. The Chengcun watershed is selected as the study area, and uncertainty analysis is carried out. The numerical simulations indicate that the performance of the proposed algorithm is significantly improved, preserving the physical meaning and reasonable range of the model parameters while significantly improving the diversity and reliability of the sampled parameter population.