Incorporated flexible load forecasting based on non-intrusive load monitoring: A TCN-based meta learning approach

Yun Zhang ^* Quanyan Shu Feng Ding Feng Liu Shuiming Jiang Wenlong Wu

• State Grid Jiangsu Electric Power Co., Ltd., Yancheng Power Supply Company, yancheng, China

Precise load forecasting of residential flexible resources with complex operation behavior pattern is pivotal for demand side management to ensure efficient energy use and stabilization of power supply and demand. To address the uncertainty and volatility characteristics within the residential flexible resources, a Temporal Convolutional Network (TCN)-based meta-learning architecture is proposed by integrating feature extraction in load disaggregation with model adaptation for improved load forecasting performance. In the preprocessing phase, Concatenated Fourier Features (CFF) are employed to emphasize periodicity in power consumption data. Subsequently, a TCN base model with an extended perception field is leveraged to capture both long-term and short-term periodic dependencies for residential load profiles. Finally, a meta-learning framework is built with a twotiered learning process through adapting the features from load disaggregation task to the flexible load forecasting task. Comprehensive evaluations using public datasets demonstrate the robustness and effectiveness of the proposed method, significantly outperforming baseline models in forecasting accuracy for flexible loads.