Probabilistic forecasting of coal consumption for power plants under deep peak shaving conditions using Informer with DDPM-based uncertainty modeling

With the integration of large-scale renewable energy sources into modern power systems, coal-fired power plants are frequently forced into deep peak-shaving operations to manage fluctuating power outputs. Under these operational conditions, coal consumption exhibits significant non-stationary characteristics, posing considerable challenges to intelligent dispatch optimization and coordinated pollutant control systems. Effective forecasting of coal consumption is thus critical. While conventional long short-term memory (LSTM) and Transformer models are capable of capturing both short-term and long-term temporal dependencies, their performance deteriorates significantly when handling highly volatile coal consumption data due to the absence of uncertainty modeling mechanisms.To address these challenges, this paper proposes a novel coal consumption forecasting method, termed DDPM-Informer, that integrates an Informer architecture with a Denoising Diffusion Probabilistic Model (DDPM). First, the Informer network effectively captures long-term dependencies using a sparse attention mechanism, enhancing computational efficiency and preliminary forecasting accuracy. Second, we innovatively introduce a DDPM-based diffusion mechanism into the latent layers of the Informer model to implicitly capture and model uncertainties embedded in the data. Unlike traditional explicit uncertainty modeling methods, the proposed DDPM mechanism requires no manual specification of modeling functions or parameters, as it implicitly models uncertainties through a latent diffusion-denoising process.Experimental validation using real operational data from a 2 × 1000 MW coal-fired power plant in Wuhan, China, demonstrates the superior accuracy of the proposed DDPM-Informer method compared with existing state-of-the-art models. The proposed method provides a new technical pathway for intelligent dispatch, combustion efficiency optimization, and emission control in coal-fired power plants under deep peak-shaving scenarios.