Characterizing the Hydraulic Connection of Cascade Reservoirs for Short-Term Generation Dispatching via Gaussian Process Regression

Abstract

The characterization of the mapping relationship (MR) between outflow of the upstream reservoir (OUR) and inflow of the downstream reservoir (IDR) in the short-term generation dispatching of cascade reservoirs (SGDCR) greatly impacts the safe and economic operation of hydropower plants. If this MR is not characterized properly, the operation process of hydropower stations will deviate from the planning dispatching schemes. Especially when the upstream reservoir (UR) undertakes peak load regulation tasks frequently and the downstream reservoir (DR) owns weak re-regulation capacity, the safety of cascade reservoirs will be seriously threatened. In this extreme system, the commonest characterizing models in SGDCR, such as the lag time (LT) model and the Muskingum model, may cause a huge deviation when simulating or forecasting the IDR. Given this dilemma, the Gaussian process regression (GPR) model, which is a representative for Bayesian regression methods, is firstly introduced to handle this MR and compared with the LT and the Muskingum model in this paper. The Three Gorges-Gezhouba cascade reservoirs (TGGCR) in China are selected as a typical case study. The performance indicators of four categories are adopted to evaluate the simulated IDR in a single period and a set of pivotal indicators are proposed to estimate the simulation dispatching in multiple periods. The results show that (1) The GPR model reduces the mean absolute deviation (MAD) about inflow of Gezhouba by 197m 3 /s and 263m 3 /s than the LT and the Muskingum model respectively; (2) The distribution characteristics of inflow deviation produced by the GPR model are most competitive. Meanwhile, the GPR model has the strongest ability when conducting the multi-period simulation dispatching and owns best applicability on the whole range of streamflow. (3) The Muskingum model is not recommended to characterize the hour-scale hydraulic connection in the extreme system of SGDCR. INDEX TERMS cascade reservoirs, short-term generation dispatching, hour-scale hydraulic connection, Gaussian process regression.

Forecast Methods for Time Series Data: A Survey

“…The x . Gaussian Process Regression (GPR) [69] is a nonparameteric model that uses Gaussian Process (GP) priors to perform regression analysis on data. The model assumptions of GPR include noise (regression residual) and Gaussian process prior.…”

Water Balance Calculation Based on Hydrodynamics in Reservoir Operation

“…Inflow calculation methods, also known as flow propagation, can be divided into two categories, hydrological methods and hydrodynamic methods. The lag time method and Muskingum method are the most commonly used hydrological methods for flow propagation [8]. The lag time method only considers the time delay, which delays the outflow from the upper reservoir for a period of time without changing its value.…”

A Sequential Structure for Water Inflow Forecasting in Coal Mines Integrating Feature Selection and Multi-Objective Optimization

“…Additionally, Li et al [8] and Goyal et al [9] developed the gaussian process regression model (GPR) and support vector regression (SVR) for water inflow prediction, respectively. Chen et al [10] applied the GPR model to handle the mapping relationship between outflow of the upstream reservoir and inflow of the downstream reservoir, where the the simulation results illustrate that the smaller mean absolute deviation can be obtained by the developed GPR model. Liu et al [11] proposed three novel hybrid wind speed forecasting models based on multi-decomposing strategy and extreme learning machine (ELM), where ELM is performed as the predictor for subseries prediction.…”

“…Additionally, the parameter setting for the above contrast models and the proposed one will be recommended in this part. Among the contrastive models, the hyper-parameters including the regularization coefficient C, kernel parameter δ and the number of hidden layer neurons within SVR and ORELM are determined by grid searching in the scope of [2 −10 , 2 10 ], [2 −8 , 2 8 ] and [10,100]. For BPNN and LSTM network, the number of hidden layer for them is set as 1, while the number of the hidden layer neurons are determined to 32 according to the trial-and-error procedure [41], [42].…”

“…Moreover, the key parameters of CEEMDAN including the maximum number of sifting iterations, standard deviation of the added white noise, and the number of realizations are preset as 5000, 500 and 0.2 following the previous investigations [43], [44]. For the proposed model, the arguments of MOGWO, such as searching agents, maximum iteration and archive size are set as 100, 50 and 100, while the boundaries of the parameters within ORELM are [0.01, 1000] and [10,100], respectively. Specifically, the decomposed components obtained by EMD, CEEMDAN and VMD are illustrated in Fig.…”