Analogue Correction Method of Errors by Combining Statistical and Dynamical Methods

Based on the atmospheric analogy principle, the inverse problem that the information of historical analogue data is utilized to estimate model errors is put forward and a method of analogue correction of errors （ACE） of model is developed in this paper. The ACE can combine effectively statistical and dynamical methods, and need not change the current numerical prediction models. The new method not only adequately utilizes dynamical achievements but also can reasonably absorb the information of a great many analogues in historical data in order to reduce model errors and improve forecast skill. Purthermore, the ACE may identify specific historical data for the solution of the inverse problem in terms of the particularity of current forecast. The qualitative analyses show that the ACE is theoretically equivalent to the principle of the previous analogue-dynamical model, but need not rebuild the complicated analogue-deviation model, so has better feasibility and operational foreground. Moreover, under the ideal situations, when numerical models or historical analogues are perfect, the forecast of the ACE would transform into the forecast of dynamical or statistical method, respectively.

Multivariate Two-stage Adaptive-stacking Prediction of Regional Integrated Energy System

To reduce environmental pollution and improve the efficiency of cascaded energy utilization, regional integrated energy system(RIES) has received extensive attention. An accurate multi-energy load prediction is significant for RIES as it enables stakeholders to make effective decisions for carbon peaking and carbon neutrality goals. To this end, this paper proposes a multivariate two-stage adaptive-stacking prediction(M2ASP) framework. First, a preprocessing module based on ensemble learning is proposed. The input data are preprocessed to provide a reliable database for M2ASP, and highly correlated input variables of multi-energy load prediction are determined. Then, the load prediction results of four predictors are adaptively combined in the first stage of M2ASP to enhance generalization ability. Predictor hyper-parameters and intermediate data sets of M2ASP are trained with a metaheuristic method named collaborative atomic chaotic search(CACS) to achieve the adaptive staking of M2ASP. Finally, a prediction correction of the peak load consumption period is conducted in the second stage of M2ASP. The case studies indicate that the proposed framework has higher prediction accuracy, generalization ability, and stability than other benchmark prediction models.