[Objective] With the continuous advancement of China’s electricity spot market reform, electricity prices experience real-time fluctuations driven by dynamic supply and demand. Concurrently, under the national "dual carbon" strategic goals, deep peak shaving of thermal power units has become normalized. This operational shift causes grid dispatch commands to exhibit intense randomness and high-frequency volatility. Traditional prediction models exhibit insufficient accuracy in predicting coal consumption and suffer from severe transient response lag when thermal power units operate under deep peak shaving conditions. To address these critical issues, this paper proposes a dual-stream fusion prediction model based on conditional variational autoencoder and extreme learning machine (CVAE-ELM), incorporating multi-scale spatiotemporal feature fusion. The primary objective is to provide power plants with a highly accurate and fast-responding tool to determine their marginal costs, thereby supporting precise bidding strategies in the electricity spot market. 

[Methods] The proposed CVAE-ELM model adopts a sophisticated dual-stream architecture to capture complex operational dynamics. First, multi-scale convolution operations and an attention mechanism are integrated into the encoder of the conditional variational autoencoder (CVAE). This configuration successfully achieves the decoupled extraction of strongly coupled operational  parameters and weakly coupled time-delay features. Second, the model establishes a dual-stream framework consisting of a deep feature stream (designated as Path A) and a raw feature stream (designated as Path B). Within this framework, residual channels are strategically utilized to capture high-frequency transient characteristics under sudden load variations. To evaluate the performance of this model, comprehensive validation was conducted using actual measured operational data collected from a real 600 MW thermal power unit under various complex working conditions. 

[Results] The experimental results demonstrate outstanding predictive performance and robust stability across diverse and changing operating conditions. The root mean square error (RMSE) of the coal consumption prediction under variable working conditions is precisely restricted to 11.047 g/(kW·h). Furthermore, the average relative error of the model is strictly controlled within a narrow margin of less than 1%. Compared with conventional single-line or single-stage prediction models, the transient response speed of the proposed dual-stream fusion model is significantly accelerated, showing an improvement of 15 % to 30 %. 

[Conclusion] This study demonstrates that the developed CVAE-ELM model achieves superior prediction accuracy when dealing with highly nonlinear relationships compared to traditional reference models, such as the standard Extreme Learning Machine (ELM) and Support Vector Regression (SVR). Even under highly complex and volatile operational conditions, the model maintains high precision and remarkable stability in coal consumption forecasting. In conclusion, this research provides an essential and precise marginal cost foundation for thermal power plants, enabling them to optimize their competitive bidding strategies within the reforming electricity spot market.