AWS-DAIE: Incremental Ensemble Short-Term Electricity Load Forecasting Based on Sample Domain Adaptation

Abstract

Short-term load forecasting is a prerequisite and basis for power system planning and operation and has received extensive attention from researchers. To address the problem of concept drift caused by changes in the distribution patterns of electricity load data, researchers have proposed regular or quantitative model update strategies to cope with the concept drift; however, this may involve a large number of invalid updates, which not only have limited improvement in model accuracy, but also insufficient model response timeliness to meet the requirements of power systems. Hence, this paper proposes a novel incremental ensemble model based on sample domain adaptation (AWS-DAIE) for adapting concept drift in a timely and accurate manner and solves the problem of inadequate training of the model due to the few concept drift samples. The main idea of AWS-DAIE is to detect concept drift on current electricity load data and train a new base predictor using Tradaboost based on cumulative weighted sampling and then dynamically adjust the weights of the ensemble model according to the performance of the model under current electricity load data. For the purposes of demonstrating the feasibility and effectiveness of the proposed AWS-DAIE algorithm, we present the experimental results of the AWS-DAIE algorithm on electricity load data from four individual households and compared with several other excellent algorithms. The experimental results demonstrated that the proposed AWS-DAIE not only can adapt to the changes of the data distribution faster, but also outperforms all compared models in terms of prediction accuracy and has good practicality.

A Machine Learning Model Ensemble for Mixed Power Load Forecasting across Multiple Time Horizons

“…The weighting of the output results of basic forecasting LSTM models in [ 68 ] is based on the similarity degree between target and identified standard values of load consumption. Two different approaches for determining the weights of multiple forecasters are followed in [ 69 , 70 ], using a novel incremental ensemble weight updating strategy and the minimum-error method, respectively. Alternatively, an extreme learning machine can be employed for combining the outputs of a pool of forecasts, as in [ 71 ].…”