Abstract: This paper proposes a bidirectional gate recurrent unit (biGRU) main steam temperature prediction model based on phase space reconstruction (PSR) to address the problem of unstable combustion state caused by peak shaving demand of thermal power units, which makes it difficult to predict frequent fluctuations in main steam temperature. First, the mutual information method is used to screen relevant variables and perform phase space reconstruction to obtain input variables. Then, the Attention Mechanism (AM) is used to determine the weight coefficients of each input variable, and the snow ablation optimization algorithm (SAO) is used to optimize the biGRU hyperparameters. The Phase Space Reconstruction- Snow Ablation Optimization-Bidirectional Gated Recurrent Unit- Attention Mechanism Steam Temperature Prediction Model (PSR-SAO-biGRU-AM Prediction Model) for the bidirectional gated cycle unit is established. Finally, the prediction model is compared and validated with models without attention mechanism, SAO algorithm, and phase space reconstruction. The results show that compared to other models, the PSR- SAO-biGRU-AM prediction model proposed in this paper has the smallest root mean square error and average absolute percentage error, and the highest prediction accuracy. It demonstrates accurate prediction capability even under severe main steam temperature fluctuations, while exhibiting excellent dynamic performance.