Hydrological Time Series Data Research Articles

Stage and discharge prediction from documentary time-lapse imagery

Imagery from fixed, ground-based cameras is rich in qualitative and quantitative information that can improve stream discharge monitoring. For instance, time-lapse imagery may be valuable for filling data gaps when sensors fail and/or during lapses in funding for monitoring programs. In this study, we used a large image archive (>40,000 images from 2012 to 2019) from a fixed, ground-based camera that is part of a documentary watershed imaging project (https://plattebasintimelapse.com/). Scalar image features were extracted from daylight images taken at one-hour intervals. The image features were fused with United States Geological Survey stage and discharge data as response variables from the site. Predictions of stage and discharge for simulated year-long data gaps (2015, 2016, and 2017 water years) were generated from Multi-layer Perceptron, Random Forest Regression, and Support Vector Regression models. A Kalman filter was applied to the predictions to remove noise. Error metrics were calculated, including Nash-Sutcliffe Efficiency (NSE) and an alternative threshold-based performance metric that accounted for seasonal runoff. NSE for the year-long gap predictions ranged from 0.63 to 0.90 for discharge and 0.47 to 0.90 for stage, with greater errors in 2016 when stream discharge during the gap period greatly exceeded discharge during the training periods. Importantly, and in contrast to gap-filling methods that do not use imagery, the high discharge conditions in 2016 could be visually (qualitatively) verified from the image data. Half-year test sets were created for 2016 to include higher discharges in the training sets, thus improving model performance. While additional machine learning algorithms and tuning parameters for selected models should be tested further, this study demonstrates the potential value of ground-based time-lapse images for filling large gaps in hydrologic time series data. Cameras dedicated for hydrologic sensing, including nighttime imagery, could further improve results.

Quantifying the natural flood management potential of leaky dams in upland catchments, Part I: A data-based modelling approach

A novel application of data-based time series methods is proposed in this study to help overcome barriers to quantifying the impacts of Natural Flood Management measures from hydrological timeseries data which hitherto have prevented accurate assessment of the effectiveness of interventions. To demonstrate the value of this method, a transfer function noise model was fitted to stage data from a three year before-after-control-impact style monitoring study of leaky dams in an upland catchment in North Yorkshire, England. Using the data-based time series method, uncertainties associated with stage data were overcome. The models were able to simulate the peaks of flood events on one stream to within ±2 cm accuracy for 95% of events recorded during the baseline monitoring period. These simulations are used in this study’s companion paper to quantify leaky dam impacts on flood peak magnitude. The level of accuracy achieved in this study provides proof of concept for application of the approach to data from other environments and natural flood management interventions, which is crucial if natural flood management is to be used as a mainstream flood risk management measure.

Early Flood Monitoring and Forecasting System Using a Hybrid Machine Learning-Based Approach

The occurrence of flash floods in urban catchments within the Mediterranean climate zone has witnessed a substantial rise due to climate change, underscoring the urgent need for early-warning systems. This paper examines the implementation of an early flood monitoring and forecasting system (EMFS) to predict the critical overflow level of a small urban stream on Lesvos Island, Greece, which has a history of severe flash flood incidents requiring rapid response. The system is supported by a network of telemetric stations that measure meteorological and hydrometric parameters in real time, with a time step accuracy of 15 min. The collected data are fed into the physical Hydrologic Engineering Center’s Hydrologic Modeling System (HEC-HMS), which simulates the stream’s discharge. Considering the HEC-HMS’s estimated outflow and other hydro-meteorological parameters, the EMFS uses long short-term memory (LSTM) neural networks to enhance the accuracy of flood prediction. In particular, LSTMs are employed to analyze the real-time data from the telemetric stations and make multi-step predictions of the critical water level. Hydrological time series data are utilized to train and validate the LSTM models for short-term leading times of 15 min, 30 min, 45 min, and 1 h. By combining the predictions obtained by the HEC-HMS with those of the LSTMs, the EMFS can produce accurate flood forecasts. The results indicate that the proposed methodology yields trustworthy behavior in enhancing the overall resilience of the area against flash floods.

New method of precipitation forecast model and validation

There is a lot of time series data in many realistic sectors with different forecast techniques over the years. However there is no unanimous conclusion on forecast techniques such as individual forecasts Autoregressive, Moving averages, Autoregressive Moving average, Autoregressive Integrated Moving average, Artificial Neural Network, Long Short Term Memory network and Auto-Regressive Conditional Heteroscedasticity / Generalized Autoregressive Conditional Heteroskedasticity and combination of forecast (simple Average of forecasts, Minimum variance method, and Regression method of the combine). The most empirical hydrological time series models do not accurately forecast the weather. This paper focuses on a comparative study of different existing individual and combination forecasts with the proposed Hybrid Stochastic Model (HSM) forecast procedure. For this we consider a hydrological time series data of the Indian subcontinent to test the proposed forecast model. As a whole in comparison to all other traditional model's contributions accuracy, the proposed model performed well, and also we examined the model's dimension reduction approach to choose an optimum number of forecast techniques to be included in the model to yield the best forecasts.

Microclimatic Variation in Miri Region (NW Borneo): Inference from Rainfall and Temperature Trends, Isotopic Signature and Air Mass Movement

Trend analysis is frequently utilised to identify the changes in meteorological and hydrologic time series data, such as rainfall and temperature. The variations in the intensity, rainfall pattern and temperature have gradually changed globally. Hence, in this study, an attempt was made to analyse the decadal rainfall and surface air temperature data to understand the microclimatic variations in the Miri coastal region of NW Borneo. A data series of records for daily total rainfall amount and daily surface temperature of 11 years from 2010 to 2021 was studied and analysed. In addition, representative rainwater and groundwater samples were collected and analysed for hydrochemical parameters and oxygen and hydrogen isotopes. A detailed literature review was carried out on rainfall patterns in Malaysia, which was used for the comparative study. Interpretation of results shows that the northeast monsoon (NEM) contributed a higher total rainfall rate with lower daily mean surface air temperature over the years compared to the southwest monsoon (SWM). The recorded data for rainfall amounts in SWM for the month of May, July, August and September were higher, particularly for the years 2010 and 2020. During NEM, a higher rainfall amount was recorded in the month of January for several years. February month has always been among the driest month in NEM, and September has been the wettest month throughout the year during SWM. The isotopic values of rainwater indicate a similar moisture source to the regional precipitation trend. Groundwater isotopes reveal the low water-rock ratio of retrograde exchange between water and primary silicate minerals in the aquifer. The moisture source of the precipitation was contributed from both oceanic and continent, affecting the rainfall intensity in this region. This study is a crucial outcome to determine the potential impacts of microclimatic variations on the rainfall patterns in the Miri coastal region.

Annual water balance and hydrological trends in the glacierised Tarfala Catchment, Sweden

Quantifying components of the hydrological cycle in glacierised catchments is important for the assessment of the temporal distribution, quantity and quality of water resources available to downstream regions, especially under a changing climate. However, this assessment requires long time series of observations, which are typically unavailable for remote catchments, such as those in mountainous areas. In this study, we leverage a unique ∼40 year time series of hydrological data recorded in the subarctic glacierised Tarfala catchment (Sweden) to explore temporal trends in the components of the catchment water balance (precipitation, runoff, change in storage, and evaporation), and to assess if water balance residuals are associated with specific hydro-climatic conditions. No significant temporal trends were found in precipitation and storage changes of the glacierised area, but significant increases were found in evaporation and summer discharge (in part attributed to glacier volume losses). The annual water balance could not be perfectly closed, and water losses were on average 112 mm y−1 larger than the water inputs over the study period. Among the water balance components, discharge contributed most to the total water balance uncertainty, and storage surplus due to antecedent meteorological conditions could explain why water losses in specific years exceeded the uncertainty bounds. It is therefore essential to consider legacy effects from previous years when applying water balance calculations in mountainous and/or glacierised catchments.

Suitable sites for groundwater development: a capture map-based approach integrated with weighted overlay analysis

Abstract A novel methodology for suitable site selection for groundwater development based on river capture, pumping cost, and groundwater potential has been proposed for better groundwater utilisation. River capture and cost map have been generated from a calibrated groundwater model, simulated with forecasted hydrological time series data. The groundwater potential has been calculated with weighted overlay analysis. These three variables have been used to classify the model domain into five zones of groundwater development by K-Means clustering. The area with lower river capture, low cost of pumping, and high groundwater potential is found to be the best location for groundwater extraction. The methodology has been applied to the lower Ain River basin, in France.

A Practical Yet Accurate Real-Time Statistical Analysis Library for Hydrologic Time-Series Big Data

Using different statistical analysis methods to examine hydrologic time-series data is the basis of accurate hydrologic status analysis. With the wide application of the Internet of Things and sensor technologies, traditional statistical analysis methods are unable to meet the demand for real-time and accurate hydrologic data analysis. The existing mainstream big-data analysis platforms lack analysis methods oriented to hydrologic data. In this context, a real-time statistical analysis library based on the new generation of big data processing engine Flink, called HydroStreamingLib, was proposed and implemented. Furthermore, in order to prove the efficiency and handiness of the proposed library, a real-time statistical analysis system of hydrologic stream data was developed based on the concepts available in the proposed library. The results showed that HydroStreamingLib provides users with an efficient, real-time statistical verification method, thus extending the application capabilities of Flink Ecology in some specific fields.

A Univariate Extreme Value Analysis and Change Point Detection of Monthly Discharge in Kali Kupang, Central Java, Indonesia

Kali Kupang plays an important role in the life of the people of Pekalongan and its surrounding areas. However, until recently, not many hydrological studies have been carried out in this area. This study presents how Extreme Value Analysis (EVA) can predict future extreme hydrological events and how a dynamic-programming-based change point detection algorithm can detect the abrupt transition in discharge events variability. Using the annual block maxima, we can predict the upper extreme discharge probability from the generalized extreme value distribution (GEVD) that best fits the data by using the Markov Chain Monte Carlo (MCMC) algorithm as a distribution fitting method. Metropolis-Hasting (MH) algorithm with 500 walkers and 2,500 samples for each walker is used to generate random samples from the prior distribution. As a result, this discharge data can be categorized as a Gumbel distribution (  = 6.818,  = 3.456, and  = 0). The recurrence intervals (RI) for this discharge data can be calculated through this distribution. The changepoint location of the annual standard deviation of this discharge data in the mid-1990s is detected by using the pruned exact linear time (PELT) algorithm. Despite some shortcomings, this study can pave the way for using data-driven algorithms, along with more traditional numerical and descriptive approaches, to analyze hydrological time-series data in Indonesia. This is crucial, considering an increasing number of hydro climatological disasters in the future as a consequence of global climate change.

Comparison of the hydrological time series modeling by the floods in river Indus of Pakistan

Today, in the field of science and technology, huge forecasting applications are used by scholars to forecast future values. Nowadays, using estimating the flood forecasting for peak flow discharges is very common for the risk assessment annually by quantitative data collections from different resources. The very famous and longest rivers of Pakistan i.e. Indus River and other rivers too like River Jhelum, River Kabul, and River Chenab are the prime sources of flooding. These rivers are the prime tributaries of the Indus River System. Pakistan's longest river, River Indus, is connected with the seven (7) gauge stations called Dams and barrages, and they are playing a vital role in the generation of electricity and also in irrigation for Pakistan. In this research paper, we calculated the flood risk for the Indus using the streamflow discharges on the daily basis. At present, Adaptive Neuro-Fuzzy Inference System (ANFIS) model is widely used to analyze these hydrological time series data. Adaptive Neuro-Fuzzy Inference Systems (ANFIS) merges the potentiality of Fuzzy Inference Systems (FIS) and Artificial Neural Networks (ANN) to work out problems of different kinds. For this purpose, we used the data for the years from 2002 to 2012 daily (6-months each year) streamflow period. In our analysis, the root means square error (RMSE) shows that the ANFIS model generated more satisfactory results than other models with minimum prediction errors. The ANFIS model is more reliable and has the feasibility of integrating the essence of a fuzzy system into the real world.1–28

Development and Application of an Urban Flood Forecasting and Warning Process to Reduce Urban Flood Damage: A Case Study of Dorim River Basin, Seoul

Early and accurate flood forecasting and warning for urban flood risk areas is an essential factor to reduce flood damage. This paper presents the urban flood forecasting and warning process to reduce damage in the main flood risk area of South Korea. This process is developed based on the rainfall-runoff model and deep learning model. A model-driven method was devised to construct the accurate physical model with combined inland-river and flood control facilities, such as pump stations and underground storages. To calibrate the rainfall-runoff model, data of gauging stations and pump stations of an urban stream in August 2020 were used, and the model result was presented as an R2 value of 0.63~0.79. Accurate flood warning criteria of the urban stream were analyzed according to the various rainfall scenarios from the model-driven method. As flood forecasting and warning in the urban stream, deep learning models, vanilla ANN, Long Short-Term Memory (LSTM), Stack-LSTM, and Bidirectional LSTM were constructed. Deep learning models using 10-min hydrological time-series data from gauging stations were trained to warn of expected flood risks based on the water level in the urban stream. A forecasting and warning method that applied the bidirectional LSTM showed an R2 value of 0.9 for the water level forecast with 30 min lead time, indicating the possibility of effective flood forecasting and warning. This case study aims to contribute to the reduction of casualties and flood damage in urban streams and accurate flood warnings in typical urban flood risk areas of South Korea. The developed urban flood forecasting and warning process can be applied effectively as a non-structural measure to mitigate urban flood damage and can be extended considering watershed characteristics.

CNN-BiLSTM water level prediction method with attention mechanism

Hydrological time series data is stochastic and complex, and the importance of its historical features is different. A single model is difficult to overcome its own limitations when dealing with hydrological time series prediction problems, and the prediction accuracy of a single model can be further improved. According to the characteristics of hydrological time series data, a CNN-BiLSTM water level prediction method with attention mechanism is proposed. In this paper, CNN extracts the spatial characteristics of water level data and BiLSTM learns the time period characteristics by combining the past and future sequence information, attention mechanism is introduced to focus the salient features in the sequence. Taking the hourly water level data of Pinghe basin in China as experimental basis, experimental result shows that this method is more accuracy than Support Vector Machine (SVM), Temporal Convolutional Neural network (TCN), and Bidirectional Long Short-Term Memory network (BiLSTM) model.

HydroMet: A New Code for Automated Objective Optimization of Hydrometeorological Thresholds for Landslide Initiation

Landslide detection and warning systems are important tools for mitigation of potential hazards in landslide prone areas. Traditionally, warning systems for shallow landslides have been informed by rainfall intensity-duration thresholds. More recent advances have introduced the concept of hydrometeorological thresholds that are informed not only by rainfall, but also by subsurface hydrological measurements. Previously, hydrometeorological thresholds have been shown to improve capabilities for forecasting shallow landslides, and they may ultimately be adapted to more generalized landslide forecasting. We present HydroMet, a code developed in Python by the U.S. Geological Survey, which allows users to guide the automated estimation of hydrometeorological thresholds for a site or area of interest, with the flexibility to select preferred threshold variables for the antecedent hydrologic conditions and the triggering meteorological conditions. Users can import hydrologic time-series data, including rainfall, soil-water content, and pore-water pressure, along with the times of known landslide occurrences, and then conduct objective optimization of warning thresholds using receiver operating characteristics. HydroMet presents many additional options, including selecting the threshold formula, the timescale of possible threshold variables, and the skill statistics used for optimization. Users can develop dual-stage thresholds for watch and warning alerts, with a lower, risk-averse threshold to avoid missed alarms and a less conservative threshold to minimize false alarms. Users may also choose to split their inventory data into calibration and evaluation subsets to independently evaluate the performance of optimized thresholds. We present output and applications of HydroMet using monitoring data from landslide-prone areas in the U.S. to demonstrate its utility and ability to produce thresholds with limited missed and false alarms for informing the next generation of reliable landslide warning systems.

Analysis of spatio-temporal trend in groundwater elevation data from arsenic affected alluvial aquifers – Case study from Murshidabad district, West Bengal, Eastern India

Fluctuation in groundwater level is a time-dependent stochastic process. It is also a function of various inflow and outflow components to and from the hydrologic system concerned. Depth to water level data are measured through a network of observation wells or hydrograph stations to ascertain the degree of fluctuation in groundwater level at the desired scale, on a long-term basis. Basically, these depths to water level data are point measurements, which can be regarded as random variables furnishing changes in groundwater storage over time. The intrinsic in-homogeneity in aquifer materials introduces variations like jumps, trends, and periodicities in such hydrologic time series data. Thus, trending results from certain gradual, natural and/or anthropogenic interventions in the hydrologic environment and analyses of their trends are imperative for assessment of groundwater level scenario in the area of interest. This in turn, is essential for strategic planning and management for exploitation of the precious groundwater resource in the same area. The area of interest in this article, i.e., Murshidabad is one of the nine arsenic affected districts of West Bengal. Here, contamination persist within, shallow, arseniferous, alluvial aquifers, which are otherwise widely exploited for irrigation purposes. According to many researchers working in this area, over-exploitation of groundwater is the root cause for the plummeting water level and the widespread arsenic contamination as well. The present study intends to detect and analyze the trends persisting in the depth to water level data measured over a period from 1996 to 2016, in Murshidabad district of West Bengal, India, amidst its complex and contrasting hydrogeologic set-up and interpret the results in terms of the hydrologic attributes of the Bengal basin as a whole. The non-parametric Mann–Kendall test and the Sen’s slope estimator have been used to identify the linear trend persisting in the time series pre- and post-monsoon groundwater elevation values. The analysis indicates statistically significant decline in water level across the study area especially during the post-monsoon season. This can be attributed to the recharge–discharge disparity within the hydrologic regime; brought through intense pumping over the study area. Its ill effect being particularly observed in the western part of river Bhagirathi. Findings of such study are crucial for assessment of dynamic groundwater resources of the district and subsequently can be utilized as a decision support tool for groundwater management at micro-level.

Wavelet based hybrid ANN-ARIMA models for meteorological drought forecasting

Drought prediction is an important subject, particularly in drought-hydrology, and has a key role in risk management, drought readiness and alleviation. Hydrological time series data consists of nonlinear features and various time scales. With this view in mind, this study has combined the strengths of the Wavelet transformation, Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Network (ANN) to test a new method of a hybrid model for their ability to accurately predict future droughts. A 30-year rainfall data from the year 1986 to 2016 for Malaysia’s Langat River Basin was analyzed. Meteorological drought indices (DI) such as the Standardized Precipitation Index (SPI) and the Standard Index of Annual Precipitation (SIAP) were used to compute historical drought events. At first, each of these computed drought time series went through a process of decomposition to be divided as low frequency and high-frequency sub-series by discrete wavelet transform (DWT). Secondly, the high and low-frequency sub-series were passed through the predictive model of ANN and ARIMA techniques, respectively. Lastly, the predicted sub-series were used to reconstruct and develop a final drought prediction model. It was found that the Wavelet-ARIMA-ANN (which named as W-2A) model outperformed the single ANN and wavelet-ANN predictive models. The ANN model developed by SPI achieved an overall correlation co-efficient R-value of 0.423, but the wavelet-based ANN model decreased in the R-value to 0.415. Finally, two different models, which were established using drought indices SPI and SIAP, and discrete wavelet transformation-based hybrid ANN-ARIMA (W-2A), have achieved improved R values of 0.914 and 0.934 respectively.

Hydrological response to climate change and human activities: A case study of Taihu Basin, China

Climate change and human activities have changed a number of characteristics of river flow in the Taihu Basin. Based on long-term time series of hydrological data from 1986 to 2015, we analyzed variability in precipitation, water stage, water diversion from the Yangtze River, and net inflow into Taihu Lake with the Mann-Kendall test. The non-stationary relationship between precipitation and water stage was first analyzed for the Taihu Basin and the Wuchengxiyu (WCXY) sub-region. The optimized regional and urban regulation schemes were explored to tackle high water stage problems through the hydrodynamic model. The results showed the following: (1) The highest, lowest, and average Taihu Lake water stages of all months had increasing trends. The total net inflow into Taihu Lake from the Huxi (HX) sub-region and the Wangting Sluice increased significantly. (2) The Taihu Lake water stage decreased much more slowly after 2002; it was steadier and higher after 2002. After the construction of Wuxi urban flood control projects, the average water stage of the inner city was 0.16–0.40 m lower than that of suburbs in the flood season, leading to the transfer of flooding in inner cities to suburbs and increasing inflow from HX into Taihu Lake. (3) The regional optimized schemes were more satisfactory in not increasing the inner city flood control burden, thereby decreasing the average water stage by 0.04–0.13 m, and the highest water stage by 0.04–0.09 m for Taihu Lake and the sub-region in the flood season. Future flood control research should set the basin as the basic unit. Decreasing diversion and drainage lines along the Yangtze River can take an active role in flood control.

Detection of Changes in Hydrological Time Series During Recent Decades

Abstract An analysis of a hydrological time-series data offers the possibility of detecting changes that have arisen due to climate change or change in land use. This paper deals with the detection of changes in the hydrological time data series. The trend analysis was applied at 58 stage-discharge gauging stations that are located throughout Slovakia, with the measurement period from 1962 to 2017. The Mann-Kendall test show a declining trends in the summer and a few rising trends in the winter in discharges. In the town of Banská Bystrica at a station on the Hron River, decades of discharges, air temperatures, and precipitation totals were analyzed. The five decades from the 1960s to the 2000s were used. The hydrological time data series were also analyzed by the Pettitt’s test, which is used to detect change points. The decadal analysis at the Banská Bystrica station shows an increase in the air temperature but insignificant changes in discharges and precipitation. Pettitt’s test identified many change points in the 1990s in the air temperature.

Inconsistencies in Yangtze River annual maxima analyses

Yangtze River annual maxima data are available for 1877–1981, for Yichang near the Three Gorges Dam. However, subsequent annual maxima seem not available outside China. Therefore, published methodologies applied to Yangtze discharge maxima cannot always be checked independently against data, despite papers appearing in international journals. The point is illustrated with respect to apparent anomalies in parameter estimation for the generalised extreme value distribution (GEV) applied to Yichang annual maxima. Specifically, Wang et al. (Stoch Environ Res Risk Assess 31: 2281–2296, 2017) list GEV parameter estimates derived from the period 1890–2000. These parameter values give underfit of all Yichang annual maxima between 1877 and 1981. In contrast, the GEV parameter estimates of Sutcliffe (J Hydrol 96: 159–171) paper give better data fits. It may be that the underfit is related to the proposed methodology, which should then be treated with caution. Alternatively, there may be just a typographical error in a listed parameter value. Whatever the cause, the point is made that it would be helpful to have complete time series of hydrological data from China to confirm published analyses.

Detecting Pattern Anomalies in Hydrological Time Series with Weighted Probabilistic Suffix Trees

Anomalous patterns are common phenomena in time series datasets. The presence of anomalous patterns in hydrological data may represent some anomalous hydrometeorological events that are significantly different from others and induce a bias in the decision-making process related to design, operation and management of water resources. Hence, it is necessary to extract those “anomalous” knowledge that can provide valuable and useful information for future hydrological analysis and forecasting from hydrological data. This paper focuses on the problem of detecting anomalous patterns from hydrological time series data, and proposes an effective and accurate anomalous pattern detection approach, TFSAX_wPST, which combines the advantages of the Trend Feature Symbolic Aggregate approximation (TFSAX) and weighted Probabilistic Suffix Tree (wPST). Experiments with different hydrological real-world time series are reported, and the results indicate that the proposed methods are fast and can correctly detect anomalous patterns for hydrological time series analysis, and thus promote the deep analysis and continuous utilization of hydrological time series data.

Water level prediction of Lake Poyang based on long short-term memory neural network

湖泊水位是维持其生态系统结构、功能和完整性的基础.鄱阳湖受流域五河和长江来水双重影响，水位变化复杂.为了准确预测鄱阳湖水位变化，采用长短时记忆神经网络方法（LSTM）构建了鄱阳湖水位预测模型.该模型以赣江、抚河、信江、饶河和修水五河入湖流量和长江干流流量作为输入条件，预测鄱阳湖湖区不同代表站（湖口、星子、都昌、吴城和康山）的水位过程.研究以1956-1980年的水文时间序列数据作为训练集，1981-2000年作为验证集，探讨了LSTM模型输入时间窗、隐藏神经元数目、初始学习率等模型参数对预测精度的影响，并确定了鄱阳湖水位预测模型的最优参数.结果表明，采用LSTM神经网络方法可基于流域五河和长江来水量历时数据合理预测鄱阳湖不同湖区的水位过程，五站水位预测的均方根误差为0.41~0.50 m，纳什效率系数和决定系数达0.96~0.98.为考察模型训练数据集对鄱阳湖水位预测结果的影响，进一步选取了随机5年（1956-1960年）的资料和5个典型水文年（1954年、1973年、1974年、1977年和1978年）的日均流量资料来训练模型.结果显示随机5年资料作为训练数据的预测精度要差于典型年水文资料训练得到的模型，尤其是洪、枯水位的预测；由于典型水文年数据量仍远低于20年的资料，故其总体预测精度要略低于采用20年资料训练的模型.建议应用这类基于数据驱动的模型时，应该尽可能多选取具有代表性的资料来训练.;Lake water level is the basis for maintaining the structure, function, and integrity of its ecosystem. The water level change of Lake Poyang is complicated as it was affected by five rivers within the basin and the Yangtze River. To accurately predict the water level change of Lake Poyang, the long short-term memory (LSTM) is used to construct the water level prediction model of Lake Poyang. The model uses the flows of the Ganjiang River, Fuhe River, Xinjiang River, Raohe River, Xiushui River and the mainstream of the Yangtze River as input conditions to predict the water level process of different representative stations in the Lake Poyang area (Hukou, Xingzi, Duchang, Wucheng and Kangshan). The hydrological time series data from 1956 to 1980 is used as the training set, and data from 1981 to 2000 was used as the verification set. The influence of model parameters such as input time window, hidden neuron nodes and initial learning rate on prediction accuracy is discussed. The optimal parameters of the Lake Poyang water level prediction model are determined. The results show that the LSTM can accurately predict the water level at different parts of Lake Poyang based on the water flow from the five rivers and the Yangtze River. The RMSE value of the five stations is 0.41-0.50 m, and the NSE and R² are 0.96-0.98. In order to investigate the impact of the model training set on the water level prediction results of Lake Poyang, the study further selects data from 5 random years (1956-1960) and 5 typical hydrological years (1954, 1973, 1974, 1977 and 1978) daily average flow data to train the model. The results show that the prediction accuracy of random 5 years data as training set is worse than that of typical annual hydrological data training, especially the prediction of flood and dry water level; since the typical hydrological data volume is still much lower than 20 years of data, the overall prediction accuracy is slightly lower than the model with 20 years of data training. Therefore, representative data should be selected as much as possible for training, when applying such a data-driven LSTM neural network model.