Capabilities of deep learning models on learning physical relationships: Case of rainfall-runoff modeling with LSTM

Advancing climate-resilient flood mitigation: Utilizing transformer-LSTM for water level forecasting at pumping stations

Proactive management of pumping stations using artificial intelligence (AI) technology is vital for effectively mitigating the impacts of flood events caused by climate change. Accurate water level forecasts are pivotal in advancing the intelligent operation of pumping stations. This study proposed a novel Transformer-LSTM model to offer accurate multi-step-ahead forecasts of the flood storage pond (FSP) and river water levels for the Zhongshan pumping station in Taipei, Taiwan. A total of 19,647 ten-minute-based datasets of pumping operation and storm sewer, FSP, and river water levels were collected between 2014 and 2020 and further divided into training (70 %), validation (10 %), and test (20 %) datasets for model construction. The results demonstrate that the proposed model dramatically outperforms benchmark models by producing more accurate and reliable water level forecasts at 10-minute (T + 1) to 60-minute (T + 6) horizons. The proposed model effectively enhances the connections between input factors through the Transformer module and increases the connectivity across consecutive time series using the LSTM module. This study reveals interconnected dynamics among pumping operation and storm sewer, FSP, and river water levels, enhancing flood management. Understanding these dynamics is crucial for effective execution of management strategies and infrastructure revitalization against climate impacts. The Transformer-LSTM model's forecasts encourage water practices, resilience, and disaster risk reduction for extreme weather events.

Improving the simulations of the hydrological model in the karst catchment by integrating the conceptual model with machine learning models

Hydrological modelling can be complex in nonhomogeneous catchments with diverse geological, climatic, and topographic conditions. In this study, an integrated conceptual model including the snow module with machine learning modelling approaches was implemented for daily rainfall-runoff modelling in mostly karst Ljubljanica catchment, Slovenia, which has heterogeneous characteristics and is potentially exposed to extreme events that make the modelling process more challenging and crucial. In this regard, the conceptual model CemaNeige Génie Rural à 6 paramètres Journalier (CemaNeige GR6J) was combined with machine learning models, namely wavelet-based support vector regression (WSVR) and wavelet-based multivariate adaptive regression spline (WMARS) to enhance modelling performance. In this study, the performance of the models was comprehensively investigated, considering their ability to forecast daily extreme runoff. Although CemaNeige GR6J yielded a very good performance, it overestimated low flows. The WSVR and WMARS models yielded poorer performance than the conceptual and hybrid models. The hybrid model approach improved the performance of the machine learning models and the conceptual model by revealing the linkage between variables and runoff in the conceptual model, which provided more accurate results for extreme flows. Accordingly, the hybrid models improved the forecasting performance of the maximum flows up to 40 % and 61 %, and minimum flows up to 73 % and 72 % compared to the CemaNeige GR6J and stand-alone machine learning models. In this regard, the hybrid model approach can enhance the daily rainfall-runoff modelling performance in nonhomogeneous and karst catchments where the hydrological process can be more complicated.

Coupling machine learning and physical modelling for predicting runoff at catchment scale

In this paper, we present an approach that combines data-driven and physical modelling for predicting the runoff occurrence and volume at catchment scale. With that aim, we first estimated the runoff volume from recorded storms aided by the Green-Ampt infiltration model. Then, we used machine learning algorithms, namely LightGBM (LGBM) and Deep Neural Network (DNN), to predict the outputs of the physical model fed on a set of atmospheric variables (relative humidity, temperature, atmospheric pressure, and wind velocity) collected before or immediately after the beginning of the storm. Results for a small urban catchment in Madrid show DNN performed better in predicting the runoff occurrence and volume. Moreover, enriching the input primary atmospheric variables with auxiliary variables (e.g., storm intensity data recorded during the first hour, or rain volume and intensity estimates obtained from auxiliary regression methods) largely increased the model performance. We show in this manuscript data-driven algorithms shaped by physical criteria can be successfully generated by allowing the data-driven algorithm learn from the output of physical models. It represents a novel approach for physics-informed data-driven algorithms shifting from common practices in hydrological modelling through machine learning.

Bayesian model averaging by combining deep learning models to improve lake water level prediction

Water level (WL) is an essential indicator of lakes and sensitive to climate change. Fluctuations of lake WL may significantly affect water supply security and ecosystem stability. Accurate prediction of lake WL is, therefore, crucial for water resource management and eco-environmental protection. In this study, three deep learning (DL) models, including long short-term memory (LSTM), the gated recurrent unit (GRU), and the temporal convolutional network (TCN), were used to predict WLs at five stations of Poyang Lake for different forecast periods (1-day ahead, 3-day ahead, and 7-day ahead). The forecast results of the three DL models were synthesized through Bayesian model averaging (BMA) to improve prediction accuracy, and Monte Carlo sampling method was used to calculated the 90 % confidence intervals to analyze the model uncertainty. All the three DL models achieved satisfactory prediction accuracy. GRU performed best in most forecast scenarios, followed by TCN and LSTM. None of the models, however, consistently provided the optimal results in all forecast scenarios. Lake WL prediction accuracy of BMA had a further improvement in metrics of NSE and R2 in 80 % of the forecast scenarios and ranked at least top two in all forecast scenarios. The uncertainty analysis showed that the containing ration (CR) values were above 84 % while the relative bandwidth (RB) maintained reliable performance over the 7-day ahead prediction. The proposed framework in the present study can realize satisfactory WL forecast accuracy while avoiding complex comparison and selection of DL models, and it can also be easily applied to the prediction of other hydrological variables.

Convolutional neural network-multi-kernel radial basis function neural network-salp swarm algorithm: a new machine learning model for predicting effluent quality parameters

A wastewater treatment plant (WWTP) is an essential part of the urban water cycle, which reduces concentration of pollutants in the river. For monitoring and control of WWTPs, researchers develop different models and systems. This study introduces a new deep learning model for predicting effluent quality parameters (EQPs) of a WWTP. A method that couples a convolutional neural network (CNN) with a novel version of radial basis function neural network (RBFNN) is proposed to simultaneously predict and estimate uncertainty of data. The multi-kernel RBFNN (MKRBFNN) uses two activation functions to improve the efficiency of the RBFNN model. The salp swarm algorithm is utilized to set the MKRBFNN and CNN parameters. The main advantage of the CNN-MKRBFNN-salp swarm algorithm (SSA) is to automatically extract features from data points. In this study, influent parameters (if) are used as inputs. Biological oxygen demand (BODif), chemical oxygen demand (CODif), total suspended solids (TSSif), volatile suspended solids (VSSif), and sediment (SEDef) are used to predict EQPs, including CODef, BODef, and TSSef. At the testing level, the Nash-Sutcliffe efficiencies of CNN-MKRBFNN-SSA are 0.98, 0.97, and 0.98 for predicting CODef, BODef, and TSSef. Results indicate that the CNN-MKRBFNN-SSA is a robust model for simulating complex phenomena.

Stepwise decomposition-integration-prediction framework for runoff forecasting considering boundary correction

Predicting river runoff accurately is of substantial significance for flood control, water resource allocation, and basin ecological dispatching. To explore the reasonable and effective application of time series decomposition in runoff forecasting, this study proposed a novel stepwise decomposition-integration-prediction considering boundary correction (SDIPBC) framework by using the stepwise decomposition sampling method and multi-input neural network. On this basis, we implemented a hybrid forecasting model combining seasonal-trend decomposition procedures based on loess (STL) with the long short-term memory (LSTM) network called STL-LSTM (SDIPBC) to estimate mid-long term river runoff. The reliability of the method was assessed using the historical runoff series of the Lianghekou and Jinping I Reservoirs in the Yalong River Basin, China, and developed several single models and hybrid models for comparative experiments. The results show that the existing decomposition-based hybrid forecasting frameworks are not suitable for practical runoff forecasting. The proposed SDIPBC framework can avoid using future information and improve the prediction accuracy of the single prediction model. For the Nash-Sutcliffe efficiency coefficient (NSE), the ten-day runoff forecasting accuracy of STL-LSTM (SDIPBC) in Lianghekou reservoir and Jinping I Reservoirs reached 0.845 and 0.862 respectively, which improved 1.81 % and 2.38 % than the single LSTM model, indicating that this is a practical and reliable decomposition-based hybrid runoff forecasting method.

A New Rainfall-Runoff Model Using Improved LSTM with Attentive Long and Short Lag-Time

It is important to improve the forecasting performance of rainfall-runoff models due to the high complexity of basin response and frequent data limitations. Recently, many studies have been carried out based on deep learning and have achieved significant performance improvements. However, their intrinsic characteristics remain unclear and have not been explored. In this paper, we pioneered the exploitation of short lag-times in rainfall-runoff modeling and measured its influence on model performance. The proposed model, long short-term memory with attentive long and short lag-time (LSTM-ALSL), simultaneously and explicitly uses new data structures, i.e., long and short lag-times, to enhance rainfall-runoff forecasting accuracy by jointly extracting better features. In addition, self-attention is employed to model the temporal dependencies within long and short lag-times to further enhance the model performance. The results indicate that LSTM-ALSL yielded superior performance at four mesoscale stations (1846~9208 km2) with humid climates (aridity index 0.77~1.16) in the U.S.A., for both peak flow and base flow, with respect to state-of-the-art counterparts.