Scenarios Generation-based Multiple Interval Prediction Method for Electricity Prices

TL;DR

This paper presents a new interval prediction method for electricity prices using a novel scenario generation approach with a specialized GAN, improving accuracy and reliability through enhanced evaluation metrics.

Contribution

It introduces a Pattern-Diversity Conditional GAN for scenario generation and proposes new evaluation metrics for interval prediction in electricity prices.

Findings

Generated scenarios are realistic and diverse.

The method achieves higher coverage probability.

Reduced average width of prediction intervals.

Abstract

This paper introduces an innovative interval prediction methodology aimed at addressing the limitations of current evaluation indicators while enhancing prediction accuracy and reliability. To achieve this, new evaluation metrics are proposed, offering a comprehensive assessment of interval prediction methods across both all-sample and single-sample scenarios. Additionally, a novel Pattern-Diversity Conditional Time-Series Generative Adversarial Network (PDCTSGAN) is developed, designed to generate realistic scenarios and support a new interval prediction framework based on scenario generation. The PDCTSGAN model incorporates unique modifications to random noise inputs, enabling the creation of pattern-diverse and realistic scenarios. These scenarios are then utilized to produce multiple interval patterns characterized by high coverage probability and reduced average width. The proposed approach is validated through detailed case studies, and the paper concludes with a discussion of future research directions to further refine interval prediction techniques.