A Multistep Prediction Model for the Vibration Trends of Hydroelectric Generator Units Based on Variational Mode Decomposition and Stochastic Configuration Networks Authors: Shaokai Yan, Fei Chen, Jiandong Yang, Zhigao Zhao Published: 2023 ABSTRACT Accurately predicting the changes in turbine vibration trends is a key part of the operational condition maintenance of hydropower units, which is of great significance for improving both the operational condition and operational efficiency of hydropower plants. In this paper, we propose a multistep prediction model for the vibration trend of a hydropower unit. This model is based on the theoretical principles of signal processing and machine learning, incorporating variational mode decomposition (VMD), stochastic configuration networks (SCNs), and the recursive strategy. Firstly, in view of the severe fluctuations of the vibration signal of the unit, this paper decomposes the unit vibration data into intrinsic mode function (IMF) components of different frequencies by VMD, which effectively alleviates the instability of the vibration trend. Secondly, an SCN model is used to predict different IMF components. Then, the predicted values of all the IMF components are superimposed to form the prediction results. Finally, according to the recursive strategy, a multistep prediction model of the HGU’s vibration trends is constructed by adding new input variables to the prediction results. This model is applied to the prediction of vibration data from different components of a unit, and the experimental results show that the proposed multistep prediction model can accurately predict the vibration trend of the unit. The proposed multistep prediction model of the vibration trends of hydropower units is of great significance in guiding power plants to adjust their control strategies to reach optimal operating efficiency. 1. Introduction As the proportion of renewable energy sources like wind and solar grows within the power production system, addressing the challenges of grid integration for these intermittent energy sources has become crucial [1,2]. Hydroelectric power, being the pioneer in power generation technologies, serves as a linchpin for peak shifting and frequency regulation, and it often complements these intermittent energy sources, ensuring the grid operates smoothly and stably [3,4,5]. Given this backdrop, it is imperative for hydroelectric units to possess a broad operational range to cater to the regulatory demands of the system. Yet their expansive operational scope and intricate structure result in a heightened risk of turbine failures, and traditional methods of planned and shutdown maintenance no longer suffice for the safe operation of hydropower plants, making condition-based maintenance using historical operation data the prevailing trend [6]. In recent years, a significant proportion of the research concerning the operational state of hydraulic turbines has been centered on fault diagnosis. Utilizing data from vibration, pressure pulsation, and other unit-specific metrics to ascertain the operating condition of turbines has emerged as a focal area of investigation, yielding a series of notable findings [7,8,9,10]. For example, Zheng et al. [8] proposed a method for identifying the flow state in the leafless zone of the draft tube based on empirical wavelet transform (EWT) and adaptive convolutional neural networks (ACNNs), which were used to analyze and identify the different operating states of the pump water wheel. Dao et al. [11] collected fault signals caused by unit sediment abrasion through acoustic vibration sensors on the turbine runner and determined the fault frequency of the signals through wavelet transform (WT), ensemble empirical mode decomposition (EEMD), and fast Fourier transform (FFT) methods, which is an important guide for the operation and maintenance of turbines which have been maintained over sediments. Therefore, developing a high-performance HGU vibration trends prediction model to accomplish accurate forecasts of HGU state trends will help power plant decision-makers build more appropriate operation and maintenance plans. These methods have a certain reference value in detecting and identifying abnormal states in hydropower units, but it is worth noting that these data-driven methods essentially belong to the category of ex post facto analysis and can only play a role in the event of equipment failure. Meanwhile, the health state of HGUs will show a trend of gradual deterioration due to the influence of the operating environment and equipment life, which will lead to the above fault diagnosis methods being unable to detect early equipment failures in time [12]. During real-world power station inspections, personnel often prioritize monitoring fluctuations in hydraulic turbine vibration and oscillation signals. If these operational metrics exceed specific thresholds, alarms are activated. The emphasis on vibration and oscillation signals as key detection indices stems from two primary reasons: first, these signals are exceptionally convenient to monitor, and second, the majority of hydraulic turbine malfunctions manifest externally as excessive vibration or oscillation [13,14,15,16]. Consequently, accurately forecasting the trend changes in hydraulic turbine vibration and oscillation has emerged as a novel research focus [17,18,19,20,21,22]. For instance, Bi et al. [18] proposed a fused metrics method based on Pearson and distance correlation to select suitable condition variables and integrated the BLSTM network to establish a prediction model that accurately forecast the vibration trend in hydraulic turbines. Similarly, Zhou et al. [21] amalgamated signal processing, feature selection, and prediction modules within a multi-objective optimization algorithm, resulting in an accurate and efficient model for predicting the vibration trends of hydropower units. The above data-driven methods have achieved good application results in the field of vibration trends prediction of HGUs. However, most of these models only involve single-step prediction and cannot predict the change of vibration trend of hydroelectric units for a long time in the future. How to construct a multistep prediction model and realize multistep prediction of unit vibration trend is the focus of this paper. In general, the multistep prediction model of vibration trend is mainly constructed based on five multistep prediction strategies: Recursive [23], Direct [24], DirRec [25], Multiple-Input–Multiple-Output (MIMO) [26], and DIRMO [27]. Among them, the recursive strategy has the advantages of simplicity and easy operation. Meanwhile, the strategy adopts the method of rolling prediction to build a multistep prediction model step by step, which is in line with the idea of HGU vibration trends prediction. For this reason, this paper constructs a multistep model for HGU vibration trends prediction by the recursive strategy. Essentially, HGU vibration trends prediction is a kind of time series prediction based on HGUs’ historical data. For the time series prediction, Fu et al. [13] generalized the vibration trends prediction of HGUs into two aspects: signal processing and data regression. Among them, recursive signal processing methods such as EMD [28,29], EEMD [30,31] and LMD are widely used. However, recursive methods like EMD have problems such as insufficient theory, serious modal aliasing, and inability to select modal components. For EEMD, it generates redundant IMFs during the decomposition process, which may not have practical significance or contribute to the analysis. This can increase the complexity of signal decomposition and add unnecessary computations. To address these problems, Dragomiretski et al. [32] proposed variational mode decomposition (VMD). Unlike recursive decomposition algorithms, such as EMD, VMD searches for variational models by iterative methods and decomposes the signal into IMF components of different frequency bands. In this article, VMD is chosen as a preprocessing method for HGU signals to alleviate the nonstationarity problem of the acquired vibration signals. In recent years, with the continuous development of machine learning and deep learning, AI models are gradually replacing traditional statistical models. For example, Fu et al. [13] proposed an HGU vibration trends prediction model based on OVMD and LSSVM, which overcame the disadvantage of the poor prediction accuracy of a single model. Xiong et al. [19] introduced a deep learning framework into the field of HGU vibration trends prediction, constructed a hybrid CNN-LSTM prediction model, and realized the effective prediction of HGU vibration trends. However, all of the above models suffer from the problem of difficult parameter settings, which seriously reduces the generalization ability of the models. A stochastic configuration network (SCN) [33], as a kind of stochastic parameter neural network, employs a model optimizer with low computational cost and high efficiency in the randomization algorithm. Therefore, it has good performance in model efficiency and accuracy. In this article, an SCN model is utilized to predict the IMF components of the VMD, and the predicted values of all IMF components are superimposed in the final prediction results. For fast and accurate multistep prediction of vibration trends in hydroelectric units, we have conducted the following studies and made the following contributions:(1)We propose a hybrid VMD and SCN-based vibration trend prediction model for hydroelectric units.(2)We introduce the recursive strategy to enhance the VMD–SCN model, resulting in a multistep prediction model for the vibration trend of hydroelectric units.(3)We apply the proposed multistep prediction model to the trend prediction of two different signal types: vibration signals and swing signals. This further validates the effectiveness and practicality of the proposed method. (1)We propose a hybrid VMD and SCN-based vibration trend prediction model for hydroelectric units. We propose a hybrid VMD and SCN-based vibration trend prediction model for hydroelectric units. (2)We introduce the recursive strategy to enhance the VMD–SCN model, resulting in a multistep prediction model for the vibration trend of hydroelectric units. We introduce the recursive strategy to enhance the VMD–SCN model, resulting in a multistep prediction model for the vibration trend of hydroelectric units. (3)We apply the proposed multistep prediction model to the trend prediction of two different signal types: vibration signals and swing signals. This further validates the effectiveness and practicality of the proposed method. We apply the proposed multistep prediction model to the trend prediction of two different signal types: vibration signals and swing signals. This further validates the effectiveness and practicality of the proposed method. The remainder of the paper is as follows: Section 2 introduces the methods and principles of the multistep prediction model. Section 3 shows the concrete flow of the proposed framework. Section 4 carries out case study research on a real Francis turbine. Finally, Section 6 obtains the conclusions of the study. 2. Methods and Principles 2.1. Variational Mode Decomposition (VMD) Aiming to solve the problems of empirical mode decomposition (EMD), such as endpoint effect, mode aliasing, and insufficient theoretical basis, Dragomiretski and Zosso [32] proposed a new adaptive signal processing method named VMD. Different from recursive decomposition algorithms such as EMD, VMD searches variational modes through iterative methods and decomposes signals into IMF components of different frequency bands. VMD has good applications in fault diagnosis [34,35], signal noise reduction [36] and parameter identification [37]. The specific calculation method of VMD is as follows: (1)Taking the minimum sum of the estimated bandwidths of each modal component as the objective function, the constrained variational problem obtained is:(1)min∑k=1K∂tδt+jπt∗uk(t)e−jωkt22s.t ∑k=1nukt=x(t) where K represents the number of intrinsic mode function (IMF) components, ∂t is the first-order partial derivative of the function with respect to time, δt is the unit impulse function, ∗ is the convolution symbol, ukt represents the k-th IMF component, ωk represents the center frequency of the k-th mode component, and xt is the original signal. Taking the minimum sum of the estimated bandwidths of each modal component as the objective function, the constrained variational problem obtained is:(1)min∑k=1K∂tδt+jπt∗uk(t)e−jωkt22s.t ∑k=1nukt=x(t) where K represents the number of intrinsic mode function (IMF) components, ∂t is the first-order partial derivative of the function with respect to time, δt is the unit impulse function, ∗ is the convolution symbol, ukt represents the k-th IMF component, ωk represents the center frequency of the k-th mode component, and xt is the original signal. (2)To simplify the calculation, the quadratic penalty factor and Lagrange multiplier are introduced to transform Equation (1) into an unconstrained problem. (2)Luk,ωk,λ=α∑k=1K∂tδt+jπt∗uk(t)e−jωkt22+x(t)−∑k=1Kuk(t)+λ(t),x(t)−∑k=1Kuk(t) where α is the quadratic penalty factor, and λt is the Lagrange multiplier. To simplify the calculation, the quadratic penalty factor and Lagrange multiplier are introduced to transform Equation (1) into an unconstrained problem. (2)Luk,ωk,λ=α∑k=1K∂tδt+jπt∗uk(t)e−jωkt22+x(t)−∑k=1Kuk(t)+λ(t),x(t)−∑k=1Kuk(t) where α is the quadratic penalty factor, and λt is the Lagrange multiplier. (3)According to the alternating direction multiplier method to find the saddle point of Equation (2), the specific process is as follows: According to the alternating direction multiplier method to find the saddle point of Equation (2), the specific process is as follows: Set the number of mode decomposition K, initialize the frequency domain u^k1, the center frequency ωk1, and the Lagrange multiplier λ^1, and then calculate the IMF uk and center frequency ωk according to Equation (3). (3)u^kn+1=x^ω−∑i≠ku^iω+λ^ω21+2αω−ωk2ωkn+1=∫0∞ωu^kωdω∫0∞u^kωdω Calculate the Lagrange multiplier λ according to Equation (4):(4)λ^n+1ω=λ^nω+ζ1x^ω−∑u^kn+1ω where ζ1 is the noise tolerance. uk, ωk and λ are iteratively updated successively until the criterion in Equation (5) is satisfied. (5)∑kukn+1−uk222uk222