Day-ahead short-term load probability density forecasting method with a decomposition-based quantile regression forest. (15th March 2020)

• Record Type: Journal Article
• Authors: He, Feifei; Zhou, Jianzhong; Mo, Li; Feng, Kuaile; Liu, Guangbiao; He, Zhongzheng
• Is Part Of: Applied energy. Volume 262(2020)
• Publication Date: 2020-03-15
• Subjects: Short-term load forecasting -- Variational mode decomposition -- Quantile Regression Forest -- Temperature and Humidity Index -- Bayesian optimization -- Tree-structured of Parzen Estimators; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2019.114396
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Proposed a decomposition-based quantile regression forest load forecasting method. The proposed method reaches n-step prediction by training n models. Temperature and humidity index is introduced as a relevant factor. Tree-structured of Parzen Estimators based Bayesian optimization is introduced. PICP, PINAW, and CWC are introduced to evaluate the interval prediction results. Abstract: Short-term load forecast (STLF) determines the power system planning for unit commitment, which is of great significance in power dispatching. A large amount of research work has been carried out on STLF. However, with the increase of load consumption and penetration of beyond the meter distributed energy generation systems, new challenges are brought to power load forecasting. Therefore, the probability density interval prediction which can more accurately reflect the uncertainty of power grid load is particularly important. In this study, a novel new day-ahead (24 h) short-term load probability density forecasting hybrid method with a decomposition-based quantile regression forest is proposed. First, the stationarity analysis is performed, and the load sequence is decomposed into several sub-models by Variational mode decomposition (VMD). Secondly, the influence of relevant factors such as weighted temperature and humidity index (WTHI) and day type is considered and extended to each sub-model sequence. Finally, a multi-step prediction strategy is proposed to predict the result of each sub-model using Quantile Regression Forest (QRF), and the prediction results are reconstructed to obtain the complete prediction probability density by Kernel density estimation (KDE). Specifically, the Bayesian optimization algorithm based on Tree-structured of Parzen Estimators (TPE) is adopted to optimize the hyperparameters. Furthermore, to verify the performance of the proposed method, the day-ahead short-term load forecasting of the proposed method and the contrast methods including decomposition-based methods and non-decomposition based methods were studied by the real load data of Henan Province, China. The probability density prediction obtained by the experiment indicates that the proposed method can acquire the narrowest prediction intervals at different confidence.