A new paradigm based on Wasserstein Generative Adversarial Network and time-series graph for integrated energy system forecasting

Abstract

With the continuous increase in the proportion of renewable energy, accurate forecasting of various tasks within integrated energy systems (IES) is becoming increasingly important. Traditional deep learning-based time series forecasting methods typically adopt loss functions such as mean squared error (MSE) to measure the difference between predicted and actual values, and optimizing neural networks via back-propagation. However, due to the significant auto-correlation present in time series data, these methods often struggle to fully capture their inherent properties. To this end, inspired by Wasserstein Generative Adversarial Networks (WGAN), this paper proposes a novel time series forecasting paradigm named IESWGAN. Specifically, the proposed method first extracts multi-frequency information and correlated feature information effectively from the raw time series through multi-stage data processing and inputs them parallel into the generator in a channel-independent approach, followed by channel-mixing learning to fully exploit the historical data. In the discriminator, we adopt sliding window technique to convert the time series data into gray-scale images resembling approximately square matrices, and employ computer vision techniques to enhance the discriminator’s ability to capture complex patterns in time series data, thereby improving forecasting accuracy through the adversarial interaction between the generator and discriminator. We selected three typical forecasting tasks (photovoltaic forecasting, wind power forecasting, and load forecasting) from Queensland’s IES to comprehensively evaluate the accuracy and generalizability of IESWGAN. The experimental results show that the mean average error (MAE) of the proposed IESWGAN on three tasks are MAEPV=0.32, MAEWind=1.13, and MAELoad=76.38 separately, which significantly achieves higher accuracy for more than 15% compared to baseline models. Moreover, ablation studies demonstrated the critical role of each component in improving forecasting accuracy, further validating the necessity of the proposed design strategies.