在电力需求不断增长的背景下，准确的短期负荷预测（Short-term Load Forecasting， STLF）技术是保证电力系统经济安全稳定运行的重要措施。采用数据驱动的方法是近年来研究STLF的主流，它通过建立环境等因素（即特征）与负荷之间的非线性映射关系取得了优异的预测性能。但这种方法存在一些局限：在特征选择时，过滤器方法难以选择出最优特征子集而降低短期负荷预测模型（Short-term Load Forecasting Model，SLFM）的预测精度，以及包装器方法时间成本大；SLFM挖掘时序依赖和周期相似性的能力有限，其预测精度受限；在应用于大数据场景时，基于人工神经网络（Artificial Neural Network，ANN）的SLFM训练时间过长。针对以上局限，本文基于配电居民负荷大数据场景，从特征选择、SLFM以及SLFM的迁移三个方面，开展数据驱动的配电居民短期负荷预测方法研究，并应用于贵阳真实的配电居民变压器上。1.针对传统特征选择不能兼顾预测效果和时间成本的问题，提出一种基于前馈长短期记忆网络的特征选择方法。该方法可以认为是包装器和过滤器方法的结合，能够全面地考虑候选特征与负荷之间的相关性，候选特征之间的冗余性和候选特征之间的交互作用选择出最优特征子集，为后续SLFM以及SLFM的迁移提供输入特征。瑞士和贵阳数据集上的实验和可视化结果表明，对于多种其他基于ANN的SLFM，该方法选择出的最优特征子集分别平均提升了其预测精度12.1%和2.02%。2.针对现有SLFM预测精度受限的问题，提出一种基于 Transformer 的SLFM（Transformer Based SLFM，TSLFM），为后续SLFM的迁移提供模型基础。瑞士和贵阳数据集上的实验和可视化结果表明，TSLFM识别出了动态时序依赖和周期相似性两种负荷特性，同时相比其他基于ANN的SLFM，其预测精度分别平均提高了12.14%和19.96%。3.针对现有基于ANN的SLFM应用大数据场景时训练时间成本大的问题，提出了一种基于序列相似性的预测模型迁移方法。该方法主要由序列关键点识别阶段、序列相似性聚类阶段以及预测模型迁移阶段组成。贵阳数据集上的实验结果表明，在满足一定预测精度的条件下，相比其他基于ANN的SLFM，TSLFM迁移的训练时间平均减少了95.96%。

Accurate Short-term Load Forecasting (STLF) technology is an important measure for power system economic, security and stable operation as the electricity demand incresed rapidly. The data-driven method has been the mainstream research on STLF in recent years, which has achieved excellent prediction performance by modeling nonlinear mapping relationships between environmental factors (i.e. features) and loads. However, some limitations still remains in this method: Firstly, during feature selection, it‘s difficult for the filter method to select the optimal feature subset, which reduces the prediction performance of the Short term Load Forecasting Model (SLFM), and the wrapper method has a high time cost; Secondly, the ability of SLFM to mine temporal dependency and periodic similarity is limited, thus its prediction performance is limited; Thirdly, when applied to big data scenarios, the training time cost of Artificial Neural Network (ANN) based SLFMs is expensive.Aiming at the above limitations, this article conducts research on data-driven short-term load forecasting method for distribution residents from three aspects: feature selection, SLFM, and SLFM transfer, which will be applied to real distribution residential transformers in Guiyang.1.Aiming at the problem that traditional feature selection can not take both prediction performance and time cost into account, a feedforward long short-term memory network based feature selection method is proposed. This method can be considered as a combination of wrapper and filter methods, which can consider the correlation between candidate features and load, the redundancy among candidate features and the interaction among candidate features to select the optimal feature subset. The chapter will provide input features for subsequent research on SLFM and its transfer. The experimental and visualization results on Switzerland and Guiyang datasets indicate that for various ANN based SLFMs, the optimal feature subset has improved their prediction performance by an average of 12.1% and 2.02%, respectively.2.Aiming at the problem that the existing SLFMs have limited prediction performance, a Transformer based SLFM (TSLFM) is proposed, which is also as a model foundation for subsequent research on SLFM transfer. The experimental and visualization results on Switzerland and Guiyang datasets show that firstly, TSLFM recognizes two load characteristics: dynamic temporal dependency and periodic similarity. Secondly, its prediction performance is improved by an average of 12.14% and 19.96% compared to other ANN based SLFMs, respectively.3.Aiming at the problem that the ANN based SLFMs have high training time cost in big data scenarios, a sequence similarity based SLFM transfer method is proposed. This method consists of key point recognition stage, sequence similarity clustering stage, and SLFM transfer stage. The experimental results on the Guiyang dataset show that, compared to other ANN based SLFMs, TSLFM transfer reduces by an average of 95.9% training time cost while maintaining an acceptable prediction accuracy.