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

Accurate forecasting of residential flexible load is imperative for effective demand-side management, ensuring efficient energy utilisation and power supply-demand stability. Conventional methods encounter challenges due to the uncertainty and volatility of residential flexible resources.This study proposes a meta-learning architecture based on Temporal Convolutional Networks (TCN). The proposed approach is comprised of three distinct stages. Firstly, preprocessing with Concatenated Fourier Features (CFF) is employed to accentuate periodicity. Secondly, a TCN base model is utilised to capture both long-term and short-term dependencies. Thirdly, a two-tiered learning process is implemented to adapt features from load disaggregation to forecasting.The efficacy of the proposed method is evaluated using public datasets, and the results demonstrate its superiority to baseline models in terms of forecasting accuracy for flexible loads. The enhanced performance of the proposed method is attributed to the integration of feature extraction and model adaptation within a meta-learning framework.Future research could explore the incorporation of contextual information to further enhance performance.