Time series prediction with input noise based on the ESN and the EM and its industrial applications
Introduction
Industrial enterprises generally rely on supervisory control and data acquisition (SCADA) systems to collect real-time data. These industrial data contain huge information about the production processes. Accurate prediction of industrial time series data can provide guidance for the subsequent optimization and scheduling for the industrial processes. Due to the non-linear, unstable and strongly coupled nature of the actual production process, it is difficult to model mechanisms with high reliability (Zhou, Yang, & Gui, 2015). As a result, a class of data-based modeling methods has recently emerged. Data-based modeling methods model the relationship between input and output variables by using data.
The machine learning models are widely used for time series prediction, such as the relevance vector machine (RVM) (Chen, Wang, Zhao, & Wang, 2019), the least square support vector machine (LS-SVM) (Guo, He, Jiang, Chu, Malekian, & Li, 2018), the Gaussian process regression (GPR) (Zhang et al., 2019), and the neural network (Al-Saati, Omran, Salman, Al-Saati, & Hashim, 2021). To deal with the nonlinearity of the time series, many model were developed, such as the self-organizing pipelined recurrent wavelet neural network (Su, Yang, & Qiao, 2023), the REMD-MMLP with temporal-window (Yang, & Fan, 2022), the compete ensemble empirical mode decomposition with adaptive noise (Feng, et al., 2022), and the double multiplicative neuron (DMN) model with modified particle swarm optimization (Pan, et al., 2021). Particularly, the recurrent neural network (RNN) showed better performance in nonlinear time series prediction due to its excellent nonlinear approximation ability (Wei et al., 2022). However, its application is greatly limited due to its characteristics of large amount of calculation, slow solution speed, and local optima. ESN, firstly proposed by Jaeger (Jaeger, 2001), is a new type of RNN (Yang et al., 2018, Xu et al., 2019). Compared with the traditional RNN, the ESN overcame the disadvantage of local optimum by simplifying the network design and training process.
Due to the complex production environment, it is inevitable to collect noisy industrial data, which will cause uncertainty in the time series prediction model (Thiagarajan and Madden, 2008, Lu et al., 2017). For the ESN, the output uncertainty was usually concerned in some literature (Li et al., 2012, Han et al., 2020). However, it is not enough to only consider the output uncertainty. Since the ESN is a three-layer neural network, the internal state uncertainty and the input uncertainty should be also considered. Recently, Sheng et al. (Sheng, Zhao, & Ying, 2012) reported an improved ESN model with the internal state noise and output noise, where a nonlinear/linear dual estimation consisting of a nonlinear Kalman filter and a linear one was used to perform the supervised learning. The experimental results showed that the proposed method was effective and robust for noisy and chaotic time series prediction. For time series prediction, the input and output are both constructed from the same series, thus, the input and output noise should be considered simultaneously when training a prediction model.
At present, the input noise was considered in many time series prediction models. McHutchon and Rasmussen (McHutchon, & Rasmussen, 2011) reported a Gaussian process regression model (NIGP) that takes into account input noise, where local linearization is performed at each input point to approximate the posterior probability distribution of the GP. Lázaro-Gredilla and Titsias (Lázaro-Gredilla & Titsias, 2011) proposed a variational heteroscedastic Gaussian process regression (VHGPR) model, where the approximate posterior probability distributions of model parameters were calculated based on the heteroscedasticity assumption. Chen et al. (Chen et al., 2020) proposed a relevance vector machine model with uncertain inputs (RVMUI). In this work, a Gaussian approximation of the marginal likelihood was reported and the posterior distribution over the model weights was approximated with the Markov chain Monte Carlo (MCMC) algorithm. However, there are few methods related to the ESN considering input noise. Given that the ESN model exhibits very good performance on modeling the nonlinearity of time series, it is necessary to develop an input noisy version of ESN for time series prediction.
Although the output uncertainty and the internal state uncertainty are considered in the existing ESN-based models, the input uncertainty is still inevitable in an ESN. In this study, an ESN model with uncertain input (ESNUI) is proposed for time series prediction, where the internal state uncertainty and the output uncertainty are also considered. The contributions are list as follows.
- (1)
Since the introduction of input noise into the ESN makes it difficult to accurately estimate the non-linear state of the dynamic repository, the state of the dynamic repository is linearized and a Gaussian approximation is made to the input variables after the non-linear mapping.
- (2)
For learning model parameters, the Expectation-Maximization (EM) algorithm is employed to update all uncertain parameters iteratively. The forward–backward algorithm is used to predict the state of the dynamical reservoir in each iteration process.
- (3)
To verify the effectiveness of this method, two standard data sets and three industrial data sets are employed respectively. The experimental results show that the proposed method performs better than the compared methods.
This paper is structured as follows. Section 2 provides the preliminaries of the ESN. Section 3 describes the proposed ESNUI model and its derivation in detail. Section 4 shows the experimental analysis of the approach. Finally, Section 5 draws a concise conclusion.
Section snippets
Preliminaries on echo state network
The ESN model includes an input layer, a hidden layer, and an output layer. The hidden layer is also called dynamical reservoir (DR). The dynamical reservoir contains a large number of sparsely connected neurons, which contains the operating state of the system and has a short-term memory function. The network structure of ESN model is shown in Fig. 1.
The state equation in an ESN is given by
The output equation is given bywhere is the input vector
Model specification
For better representing the derivation of the proposed model, this paper uses the following simplified ESN formulations in (3) and (4). The state equation and output equation of the ESN model can be expressed as:
andwhere is the input vector without noise at time t.
In order to calculate the posterior probability of the parameters, the Gaussian likelihood function of the model is given by
Since the activation
Experimental analysis
To verify the effectiveness of the proposed ESNUI, the experiments are conducted with the proposed method and several other methods in this section. The comparison results show that the proposed method performs well in prediction accuracy. The mean absolute percent error (MAPE) and root mean square error (RMSE) are adopted as the criteria for the prediction performance. The smaller the value is, the higher the prediction accuracy becomes. The calculation formulas of MAPE and RMSE are as follows:
Conclusions
In this study, an ESNUI model is proposed to address the input uncertainty issue. After performing the linearization procedure for the state equation by the EKF, the forward-backward algorithm is used for state estimation. The model parameters are updated iteratively through the EM algorithm. Finally, the MSO data set, BFG consumption flow data of hot blast stove, BFG tank data, and BFG consumption flow data of coke oven are adopted to verify the effectiveness of the proposed method. The
CRediT authorship contribution statement
Ying Liu: Writing – original draft. Long Chen: Conceptualization, Methodology. Yunchong Li: Software, Validation. Jun Zhao: Writing – review & editing. Wei Wang: Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work described in this paper was supported by the National Key R&D Program of China under Grant 2017YFA0700300, the National Natural Sciences Foundation of China under Grant 61873048, 62003072, 62125302, and the Fundamental Research Funds for the Central Universities of China under Grant DUT20RC(3)013.
Ying Liu received the Ph. D. degree in control theory and control engineering at Dalian University of Technology, Dalian, China, in 2010. She is currently an associate professor at Dalian University of Technology. Her research interests cover integrated production planning and scheduling problem, system simulation modeling, intelligent optimization, and artificial neural network.
References (25)
- et al.
Multi-step-ahead solar output time series prediction with gate recurrent unit neural network using data decomposition and cooperation search algorithm
Energy
(2022) - et al.
Time-series interval prediction under uncertainty using modified double multiplicative neuron network
Expert Systems With Applications
(2021) - et al.
Prediction for noisy nonlinear time series by echo state network based on dual estimation
Neurocomputing
(2012) - et al.
Self-organizing pipelined recurrent wavelet neural network for time series prediction
Expert Systems With Applications
(2023) - et al.
Design of polynomial echo state networks for time series prediction
Neurocomputing
(2018) - et al.
Efficient and robust time series prediction model based on REMD-MMLP with temporal-window
Expert Systems With Applications
(2022) - et al.
Wind speed prediction method using shared weight long short-term memory network and Gaussian process regression
Applied Energy
(2019) - et al.
Statistical modeling of monthly streamflow using time series and artificial neural network models: Hindiya Barrage as a case study
Water Practice and Technology
(2021) - et al.
Relevance vector machines-based time series prediction for incomplete training dataset: Two comparative approaches
IEEE Transactions on Cybernetics
(2019) - et al.
A gaussian approximation of marginal likelihood in relevance vector machine for industrial data with input noise
IEEE Transactions on Instrumentation and Measurement
(2020)
An improved LSSVM model for intelligent prediction of the daily water level
Energies
An improved fruit fly algorithm-unscented Kalman filter-echo state network method for time series prediction of the network traffic data with noises
Transactions of the Institute of Measurement and Control
Cited by (1)
Fault diagnosis of power-shift system in continuously variable transmission tractors based on improved echo state network
2023, Engineering Applications of Artificial Intelligence
Ying Liu received the Ph. D. degree in control theory and control engineering at Dalian University of Technology, Dalian, China, in 2010. She is currently an associate professor at Dalian University of Technology. Her research interests cover integrated production planning and scheduling problem, system simulation modeling, intelligent optimization, and artificial neural network.
Long Chen received the B.S. degree in mechanical engineering from Shijiazhang Tiedao University, Shijiazhang, China, and the Ph.D. degree in Dalian University of Technology, Dalian, China, in 2012 and 2019, respectively. His research interests include industrial production scheduling and machine learning.
Yunchong Li received the B.S. degree in automation at Dalian University of Technology, Dalian, China, in 2018, and now he is a M.S. student in Dalian University of Technology, Dalian, China. His research interests include industrial production scheduling and machine learning.
Jun Zhao received the B.Sc. degree in control theory from Dalian Jiaotong University, Dalian, China, and the Ph.D. degree in engineering from Dalian University of Technology, Dalian, China, in 2003 and 2008, respectively. He is currently a Professor at the School of Control Sciences and Engineering, Dalian University of Technology. His research interests include industrial production scheduling, computer integrated manufacturing, intelligent optimization, and machine learning.
Wei Wang received the B.S., M.S., and Ph.D. degrees from Northeastern University, Evanston, IL, USA, in 1982, 1986, and 1988, respectively, all in industrial automation. He is currently a Professor with School of Control Sciences and Engineering, Dalian University of Technology, Dalian, China. His current research interests include adaptive control, computer integrated manufacturing, and computer control of industrial process.