Dynamic Multi-Objective Optimization Time Series Ensemble Prediction Framework Based on Correlation Type Detection

Dynamic multi-objective optimization problems (DMOPs) exhibit variations in constraints, decision parameter quantities, and the number of objectives over time, leading to changes in their optimal solutions. Prediction-based methods currently stand as the prevailing paradigm for handling DMOPs. Implicit correlations among solutions ob-tained at consecutive time steps while solving DMOPs con-tinuously can guide predictions for future time increments in the time series. Most existing methods rely on a single linear or nonlinear model to capture the correlations of historical optimal solutions. However, a linear model struggles with the nonlinear scenario, and using a nonlinear model to fit the linear relationship of the historical optimal solutions increases the training time because the nonlinear model has more parameters. This paper proposes a novel framework for time series ensemble prediction based on correlation type detection TSEPFCTD. The framework leverages correlation type detection strategy to identify the autocorrelation types in the time series of historical solutions. Subsequently, distinct prediction models are applied for different types, enhancing prediction accuracy. We implement this prediction framework on MOEA/D-DE, constructing a novel algorithm for solving DMOPs. Experimental results on a series of test suites demonstrate the effectiveness of our algorithm in providing robust solution outcomes.