IEEE Transactions on Industry Applications · 2024 · 42 citations · 21 references
ComputingData ScienceForecastingMachine LearningTime Series AnalysisDeep LearningEnsemble ForecastingEnsemble MethodsLoad ForecastingTime Series ForecastingTime Series PredictionDeep Learning For ForecastingDeep Learning For Time Series ForecastingDeep Learning-based ForecastingForecasting AccuracyHybrid Load ForecastingLoad DecompositionSeasonal Trend DecompositionTemporal Convolutional Networks
Accurate load forecasting is essential for the operational management of distribution network. This paper proposes a hybrid ensemble deep learning (HEDL) load forecasting framework. Two-stage load decomposition is based on multiple seasonal-trend decomposition using loess (MSTL) and variational mode decomposition (VMD), which simplifies the observed load sequence with complex variation patterns on multi-timescales and multi-frequencies. Multidimensional feature matrices considering component variability are constructed based on the maximum information coefficient (MIC) and the sliding window. The component forecasting models based on temporal convolutional network (TCN) are built, which fully capture the long-term and short-term dependencies of loads through distinct receptive fields. A weighted ensemble method is utilized to obtain more accurate forecasting results for residual components with high uncertainties. The final load forecasting result is then ensembled by summing up the multiple component forecasting results. The results of case studies demonstrate the effectiveness of the proposed HEDL. Comparative experiments with five benchmark models and two advanced frameworks have verified the superiority of the proposed HEDL framework in load forecasting accuracy, generalizability and robustness.
21