Industrial Time Series Prediction Based On Adaptive Variational Mode Decomposition

A large amount of time series data will be collected in the industrial production process.Accurate prediction of key process variables can provide an important reference for the scheduling and decision-making of industrial systems.However,industrial time series data generally have the characteristics of strong nonlinearity,non-stationarity and high noise.Traditional single-model methods are difficult to achieve effective predictions.Therefore,current research focuses on combined models based on signal decomposition.Variational modal decomposition,as a quasi-orthogonal and completely non-recursive decomposition method,effectively solves the problem of modal aliasing and has stronger anti-noise ability.Aiming at the shortcomings of the current method,such as time-consuming and insufficient input features of the combined model,this paper proposes a combined forecasting model based on adaptive variational modal decomposition and long-short-term memory network and applies it to industrial time series point forecasting and interval forecasting.Aiming at the point prediction problem of industrial time series,in order to solve the defect that the current method which does not consider the overall prediction error of the combined model when optimizing the variational modal decomposition parameters,a Bayesian parameter optimization method with mesh refinement is proposed.In this method,the objective function is to minimize the model error of the low-frequency trend and some high-frequency signals after decomposition,and the feasible region is divided into several small areas to reduce the optimization space.Secondly,in view of the poor prediction effect of the sub-model at the extreme point,an envelope model is proposed to predict the envelope of the sub-sequence at the current point to provide a reference for the amplitude,which can improve the prediction accuracy for sub-sequences.For the interval prediction problem of industrial time series,the current model training based on multi-objective optimization algorithm is time-consuming and difficult to implement.Based on the proposed point prediction model,this paper designed a loss function that includes interval coverage and interval width.On this basis,aiming at the shortcoming that the interval coverage is difficult to guarantee,an iterative algorithm for adaptively constructing the upper and lower bounds of the interval supervision information is proposed,and the convergence of the algorithm is guaranteed by presetting a prediction interval nominal confidence.To verify the validity of the method,this paper uses time series data of industrial simulation.In the point prediction experiment,the root mean square error and average percentage error are used as evaluation indicators to verify the accuracy of the method proposed in this paper in the short-term and long-term predictions.Aiming at the interval prediction experiment,a comparative experiment was designed in terms of training method and model structure,and the root mean square error and coverage width criterion were used as evaluation indicators to evaluate the interval prediction performance of the proposed method.Experimental results show that the proposed method in this paper is superior to other methods in terms of prediction accuracy and interval prediction performance.