GMINN: A Generative Moving Interactive Neural Network for Enhanced Short-Term Load Forecasting in Modern Electricity Markets

Choujun Zhan, Du Yin, Yingshan Shen, Tianyong Hao

Abstract: Short-term load forecasting is crucial for modern electricity markets. However, it is also a challenging task due to the overfitting issue of many existing models and the influence of various factors on electricity demand, such as seasons, weather and prices. To address this problem, we propose a novel short-term load forecasting framework, named Generative and Moving Interactive Neural Networks, that integrates Mixup, Moving Average Filter, Sample Convolution and Interaction Network (SCINet) and Genetic Algorithm. Firstly, a data generation component applies Mixup to augment the training dataset, reduce the overfitting issue, and enhance model generalization. Then, a decomposition component uses MAF to decompose the load data samples into trend and residual components, each representing a more predictable underlying pattern. The decomposition also prevents data leakage from future samples. Finally, a forecasting component employs SCINet to downsample and encode the trend and residual components at multiple temporal scales, capturing both long-term and short-term dependencies. A fully connected layer then decodes the encoded features to produce the load forecast. In particular, the framework uses a Genetic Algorithm to optimize its hyper-parameters automatically, addressing the issue of parameter sensitivity in deep learning networks. We test the proposed model on four real datasets from the U.S. electricity market and compare it with eight classical machine learning models and four state-of-the-art series forecasting models. The results demonstrate that our model outperforms all the baseline models in three evaluation metrics. Specifically, in terms of mean absolute percentage error (MAPE), our model achieves an average improvement of 8.7% across the four datasets compared to the best baseline score.