Advanced data-analytics for masked-load forecasting in distribution networks with high-level distributed energy resources

• Main Author: Zhou, Ziyan
• Other Authors: Xu Yan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/182404
• Institution: Nanyang Technological University

Accurate load forecasting provides statistical information to power dispatch. With distributed energy resources (DERs) creating new opportunities in the pursuing of Net Zero Emission, load forecasting is faced with new challenges. Firstly, traditional load forecasting has shifted to present masked-load forecasting (MLF), since load is masked by DERs. Secondly, DERs introduces high level of uncertainty which increase the difficulty in masked-load forecasting. Traditional load forecasting has been studied extensively, and the survey of traditional load forecasting is abundant. Owing to the excellent ability of NN in dealing with complex data, it becomes a popular method in these studies. However, traditional methods on load forecasting could not be compatible with current MLF owing to the real scenarios of MLF where masked-load information may be limited, or it is difficult to select the most related factors when considering MLF under external environment impact. To solve above threats, this Ph.D. research aims to develop several advanced data-analytics methods. These promising advanced data-analytics technologies, include transfer learning (TL), federated learning and causality learning, etc. Corresponding methods will be developed to cater to the practical scenarios of MLF. The main contributions through this Ph.D. research are as follows: Firstly, aiming to forecast masked-load with behind-the-meter DERs, transfer learning-based methods are proposed in this work. Abundant history unmasked-load which is native demand and limited present mask-load data are involved in NN construction. Model is trained by transferring common knowledge from unmasked-load to present masked-load data. According to different transferring mechanisms, several transfer-learning based methods have been proposed and tested which are Domain Invariant Neural Network (DINN), Maximum Mean Discrepancy Neural Network with Bayesian optimizations (MMD-NNb), and an ensembled NN for interval forecasting named as (DSINN). Secondly, considering a practical scenario where DERs are possessed by respective skate holders, this chapter proposed a federated learning-based approach Vertical Federated Learning-based Masked-Load Forecasting (VFL-MLF). Federated learning enables a collaborative model training where DERs data could be comprehensively utilized in masked-load forecasting while DERs’ raw data is preserved at local skate holder. Thirdly, DERs data is a pivotal part of masked-load, so is other data such as weather-related information. With relatively few works have paid attention to input feature selection in masked-load forecasting, this chapter proposed a Causality-based data-driven method for Residential Masked-Load Forecasting (CaRMLF) to select input feature based on causality relationship between meteorological variables and masked-load. All the methods have been tested on synthesis dataset and real-world dataset. Compared with related benchmark models, their effectiveness have been validated in terms of accuracy or privacy issue.