Probabilistic-Adversarial Networks (PANs): A Probability-Informed Deep Learning Approach for Electrical Demand Forecasting

Abstract

Time series forecasting has significantly advanced with the rise of deep learning models. However, recent methods often ignore the role of probabilistic state transitions in capturing temporal uncertainty. This paper introduces a novel forecasting framework based on a Probability-Adversarial Network (PAN), which leverages an adversarial architecture to embed probabilistic dynamics into neural networks. By incorporating probabilistic reasoning into a generator and discriminator architecture, the proposed PAN enhances forecasting accuracy and improves the model’s ability to handle uncertainty in sequential data. The framework is evaluated on electrical demand time series from a large-load, highly variable data center, and an average-size load, a more predictable school. Results demonstrate improved performance over probability-agnostic deep learning baselines for both load types, highlighting PAN’s versatility in addressing diverse forecasting challenges.