<p>Daily-scale carbon dioxide emission forecasting provides critical support for short-term carbon monitoring and policy evaluation. However, emission time series are often accompanied by strong noise and non-stationarity, and the response of emissions to exogenous drivers such as temperature and holidays exhibits evident scenario-dependent characteristics. These factors lead to lag or amplitude bias in existing methods. To address this, this paper proposes a novel multivariate daily-scale CO₂ emission forecasting framework, PhyQ-Mamba, driven by physical queries. This framework comprises four key components: First, variational mode decomposition (VMD) and wavelet threshold denoising (WTD) are used for data preprocessing. Next, a dual-stream Mamba backbone network is employed to model both the total emission evolution and changes in sectoral structure proportions. Furthermore, a physics-query cross attention (PQCA) mechanism is introduced, which maps exogenous variables as query signals to dynamically retrieve and fuse historical patterns relevant to the current scenario. Finally, a hierarchical consistency parameterization and physical information regularization strategy is adopted to ensure structural consistency in sectoral predictions and guide the model to learn reasonable temperature-driven and temporal evolution patterns. Experimental results show that, compared to traditional methods, the proposed model reduces root mean squared error (RMSE) by 52.72%, demonstrating significant improvements in accuracy. Ablation studies further reveal the synergistic effects of different modules within the model. In conclusion, the proposed model offers an effective solution for high-accuracy and interpretable daily-scale carbon emission forecasting.</p>