An interval decomposition-ensemble approach with data-characteristic-driven reconstruction for short-term load forecasting

•A decomposition–ensemble model is proposed for interval-valued load forecasting.•Multivariate multiscale permutation entropy technique is utilized to perform complexity analysis on decomposed interval-valued components for capturing internal features and reconstructing the components.•An automatic Bayesian optimization algorithm based on the Tree-structured Parzen Estimator algorithm is used in hyperparameter optimization.•Empirical results verify that the proposed approach outperforms other benchmarks under study. Short-term load forecasting is crucial for power demand-side management and the planning of the power system. Considering the necessity of interval-valued time series modeling and forecasting for the power system, this study proposes an interval decomposition-reconstruction-ensemble learning approach to forecast interval-valued load, in terms of the concept of “divide and conquer”. First, bivariate empirical mode decomposition is applied to decompose the original interval-valued data into a finite number of bivariate modal components for extracting and identifying the fluctuation characteristics of data. Second, based on the complexity analysis of each bivariate modal component by multivariate multiscale permutation entropy, the components were reconstructed for capturing inner factors and reduce the accumulation of estimation errors. Third, long short-term memory is utilized to synchronously forecast the upper and the lower bounds of each bivariate component and optimized by the Bayesian optimization algorithm. Finally, generating the aggregated interval-valued output by ensemble the forecasting results of the upper and lower bounds of each component severally. The electric load of five states in Australia is used for verification, and the empirical results show that the forecasting accuracy of our proposed learning approach is significantly superior to single models and the decomposition-ensemble models without reconstruction. This indicates that our proposed learning approach appears to be a promising alternative for interval load forecasting.