Accurate and reliable runoff forecasting plays an increasingly important role in the optimal management of water resources. To improve the prediction accuracy, a hybrid model based on variational mode decomposition (VMD) and deep neural networks (DNN), referred to as VMD-DNN, is proposed to perform daily runoff forecasting. First, VMD is applied to decompose the original runoff series into multiple intrinsic mode functions (IMFs), each with a relatively local frequency range. Second, predicted models of decomposed IMFs are established by learning the deep feature values of the DNN. Finally, the ensemble forecasting result is formulated by summing the prediction sub-results of the modelled IMFs. The proposed model is demonstrated using daily runoff series data from the Zhangjiashan Hydrological Station in Jing River, China. To fully illustrate the feasibility and superiority of this approach, the VMD-DNN hybrid model was compared with EMD-DNN, EEMD-DNN, and multi-scale feature extraction -based VMD-DNN, EMD-DNN and EEMD-DNN. The results reveal that the proposed hybrid VMD-DNN model produces the best performance based on the Nash-Sutcliffe efficiency (NSE?=?0.95), root mean square error (RMSE?=?9.92) and mean absolute error (MAE?=?3.82) values. Thus the proposed hybrid VMD-DNN model is a promising new method for daily runoff forecasting.
相似文献Rainfall forecast is critical to the management and allocation of water resources. Deep learning is used to predict rainfall time series with high temporal and spatial variability. Discrete wavelet transform (DWT), long-short term memory (LSTM) and dilated causal convolutional neural network (DCCNN) is integrated to build a hybrid model (DWT-CLSTM-DCCNN). Two methods of sample construction are used to train DWT-CLSTM-DCCNN and their effects on the model performance are analyzed. LSTM and DCCNN are built as benchmark models. The forecasting performance of DWT-CLSTM-DCCNN on monthly rainfall data from four major cities in China is evaluated. The results of DWT-CLSTM-DCCNN are compared with those of benchmark models in terms of the mean absolute error (MAE), root mean squared error (RMSE) and Nash-Sutcliffe efficiency (NSE) as well as the forecasting curves. The results show that DWT-CLSTM-DCCNN outperforms the benchmark models in model accuracy and peak and mutational rainfall capture.
相似文献Various time series forecasting methods have been successfully applied for the water-stage forecasting problem. Graphical time series models are a class of multivariate time series to model the spatio-temporal dependencies between the sensors. Constructing graph-based models involve data pre-processing and correlation analysis to capture the dynamics of different water flow scenarios, which is not scalable for a large network of sensors. This paper presents a novel approach to model spatio-temporal dependencies across river network stations using a partial correlation graph. We also provide a method to enrich this partial correlation graph by eliminating the spurious correlations. We demonstrate the utility of enriched partial correlation graphs in multivariate forecasting for various scenarios and state-of-the-art multivariate forecasting models. We observe that the forecasting techniques that use information from the enriched partial correlation graph outperform standard time series forecasting approaches for river network forecasting.
相似文献Predicting urban water demand is important in rationalizing water allocation and building smart cities. Influenced by multifarious factors, water demand is with high-frequency noise and complex patterns. It is difficult for a single learner to predict the nonlinear water demand time series. Therefore, ensemble learning is introduced in this work to predict water demand. A model (Word-embedded Temporal Feature Network, WE-TFN) for predicting water demand influenced by complex factors is proposed as a base learner. Besides, the seasonal time series model and the Principal Component Analysis and Temporal Convolutional Network (PCA-TCN) are combined with WE-TFN for ensemble learning. Based on the water demand data set provided by the Shenzhen Open Data Innovation Contest (SODIC), WE-TFN is compared with some typical models. The experimental results show that WE-TFN performs well in fitting local extreme values and predicting volatility. The ensemble learning method declines by approximately 68.73% on average on the Root Mean Square Error (RMSE) compared with a single base learner. Overall, WE-TFN and the ensemble learning method outperform baselines and perform well in water demand prediction.
相似文献
The protection of high quality fresh water in times of global climate changes is of tremendous importance since it is the key factor of local demographic and economic development. One such fresh water source is Vrana Lake, located on the completely karstified Island of Cres in Croatia. Over the last few decades a severe and dangerous decrease of the lake level has been documented. In order to develop a reliable lake level prediction, the application of the artificial neural networks (ANN) was used for the first time. The paper proposes time-series forecasting models based on the monthly measurements of the lake level during the last 38 years, capable to predict 6 or 12 months ahead. In order to gain the best possible model performance, the forecasting models were built using two types of ANN: the Long-Short Term Memory (LSTM) recurrent neural network (RNN), and the feed forward neural network (FFNN). Instead of classic lagged data set, the proposed models were trained with the set of sequences with different length created from the time series data. The models were trained with the same set of the training parameters in order to establish the same conditions for the performance analysis. Based on root mean squared error (RMSE) and correlation coefficient (R) the performance analysis shown that both model types can achieve satisfactory results. The analysis also revealed that regardless of the model types, they outperform classic ANN models based on datasets with fixed number of features and one month the prediction period. Analysis also revealed that the proposed models outperform classic time series forecasting models based on ARIMA and other similar methods .
相似文献Accurate and consistent annual runoff prediction in a region is a hot topic in management, optimization, and monitoring of water resources. A novel prediction model (ESMD-SE-WPD-LSTM) is presented in this study. Firstly, extreme-point symmetric mode decomposition (ESMD) is used to produce several intrinsic mode functions (IMF) and a residual (Res) by decomposing the original runoff series. Secondly, sample entropy (SE) method is employed to measure the complexity of each IMF. Thirdly, wavelet packet decomposition (WPD) is adopted to further decompose the IMF with the maximum SE into several appropriate components. Then long short-term memory (LSTM) model, a deep learning algorithm based recurrent approach, is employed to predict all components. Finally, forecasting results of all components are aggregated to generate the final prediction. The proposed model, which is applied to seven annual series from different areas in China, is evaluated based on four evaluation indexes (R, MAE, MAPE and RMSE). Results indicate that ESMD-SE-WPD-LSTM outperforms other benchmark models in terms of four evaluation indexes. Hence the proposed model can provide higher accuracy and consistency for annual runoff prediction, rendering it an efficient instrument for scientific management and planning of water resources.
相似文献Daily inflow forecasts provide important decision support for the operations and management of reservoirs. Accurate and reliable forecasting plays an important role in the optimal management of water resources. Numerous studies have shown that decomposition integration models have good prediction capacity. Considering the nonlinearity and unsteady state of daily incoming flow data, a hybrid model of adaptive variational mode decomposition (VMD) and bidirectional long- and short-term memory (Bi-LSTM) based on energy entropy was developed for daily inflow forecast. The model was analyzed using the mean absolute error (MAE), the root means square error (RMSE), Nash–Sutcliffe efficiency coefficient (NSE), and correlation coefficient (r). A historical daily inflow series of the Baozhusi Hydropower Station, China, is investigated by the proposed VMD-BiLSTM with hybrid models. For comparison, BP, GRNN, ELMAN, SVR, LSTM, Bi-LSTM, EMD-LSTM, and VMD-LSTM, were adopted and analyzed for evaluation and analyzed. We found that the proposed model, with MAE?=?38.965, RMSE?=?64.783, and NSE?=?95.7%, was superior to the other models. Therefore, the hybrid model is robust and efficient for forecasting highly nonstationary and nonlinear streamflow. It can be used as the preferred data-driven tool to predict the daily inflow flow, which can ensure the safe operation of hydropower stations in reservoirs. As an interdisciplinary field spanning both machine learning and hydrology, daily inflow forecasting can become an important breakthrough in the application of deep learning to hydrology.
相似文献