An adaptive forecasting method for the aggregated load with pattern matching

Abstract

Electrical load forecasting plays a vital role in the operation of power system. In this paper, a novel adaptive short-term load forecasting method for the aggregated load is built. The proposed method consists of two stages: load forecast model preparation stage and adaptive load forecast model selection stage. In the first stage, based on historical load data of all consumers, the typical monthly load patterns are firstly identified in an optimal fashion with the aid of the cosine similarity. Then, for each identified monthly load pattern, a stacking ensemble learning method is proposed to train the load forecasting model. In the second stage, according to the similarity between individual load data of the latest month and the identified monthly load pattern, all the consumers are firstly classified into different groups where each group corresponds to a particular load pattern. Then, for each group, the corresponding trained load forecasting model is employed for short-term load forecast and the final forecast of the aggregated load is calculated as a simple aggregation of the produced load forecast for each group of consumers. Case studies conducted on open dataset show that, compared with the single forecasting model, the proposed adaptive load forecasting method can effectively improve the load forecasting accuracy.