Probabilistic Forecasting Method for Offshore Wind Farm Cluster under Typhoon Conditions: a Score-Based Conditional Diffusion Model

TL;DR

This paper introduces a novel score-based conditional diffusion model for probabilistic offshore wind power forecasting during typhoons, effectively handling data scarcity and stochasticity to improve prediction accuracy.

Contribution

The study develops a new probabilistic forecasting method using a score-based diffusion model combined with knowledge graph embeddings for typhoon path data.

Findings

Outperforms baseline models in probabilistic metrics during typhoon conditions

Effectively captures uncertainty in offshore wind power forecasts

Demonstrates improved accuracy on real-world offshore wind farm data

Abstract

Offshore wind power (OWP) exhibits significant fluctuations under typhoon conditions, posing substantial challenges to the secure operation of power systems. Accurate forecasting of OWP is therefore essential. However, the inherent scarcity of historical typhoon data and stochasticity of OWP render traditional point forecasting methods particularly difficult and inadequate. To address this challenge and provide grid operators with the comprehensive information necessary for decision-making, this study proposes a score-based conditional diffusion model (SCDM) for probabilistic forecasting of OWP during typhoon events. First, a knowledge graph algorithm is employed to embed historical typhoon paths as vectors. Then, a deterministic network is constructed to predict the wind power under typhoon conditions based on these vector embeddings. Finally, to better characterize prediction errors, a denoising network is developed. At the core of this approach is a mean-reverting stochastic differential equation (SDE), which transforms complex error distributions into a standard Gaussian, enabling the sampling of forecasting errors using a reverse-time SDE. The probabilistic forecasting results are reconstructed by combining deterministic forecasts with sampled errors. The proposed method is evaluated using real-world data from a cluster of 9 offshore wind farms. Results demonstrate that under typhoon conditions, our approach outperforms baseline models for both deterministic and probabilistic metrics, verifying the effectiveness of the approach.