Abstract: To address the control challenges of nonlinearity, large time lag and strong coupling of molteniron temperature in blast furnace ironmaking, as well as the industrial limitations of traditional methods including separated prediction and regulation processes, highly subjective rule design, and insufficient model interpretability, this paper investigates an intelligent and explainable regulation method for molten iron temperature. The purpose is to construct an integrated closed-loop architecture of high-precision prediction and scientific regulation, improve the stability and target hit rate of molten iron temperature control, and provide technical support for intelligent and stable production of blast furnaces. Twenty two process parameters collected from the continuous production of a blast furnace from August 2024 to February 2025 were selected as research objects. DBSCAN+LOF two-level outlier detection and random forest missing value imputation were adopted for data preprocessing. Ten core features with both temporal correlation and process coupling characteristics were determined via Spearman correlation analysis and multi-method fusion feature selection. A TSMixer time series prediction model was constructed, and its time mixing layer and feature mixing layer captured the long-term time lag characteristics and parameter coupling relationships of production data, respectively. SHAP (SHapley Additive exPlanations) interpretable analysis was applied to quantify feature contributions, and four online adjustable core variables including oxygen-enriched flow rate, fuel ratio, hot blast temperature and blast kinetic energy were screened out. A data-driven fuzzy controller was designed with molten iron temperature deviation and its change rate as inputs. The controller outputted the incremental values of the four control variables, and process execution constraints were imposed to form executable regulation instructions for field application. Experimental results show that the goodness-of-fit R² of the TSMixer model reaches 88%, and the single-step inference time is only 12 ms, which delivers superior performance in both prediction accuracy and real-time capability. The proposed regulation architecture increases the target hit rate of molten iron temperature control to 92%, and narrows the temperature fluctuation range from -15 ℃ to 15 ℃ to within -5 ℃ to 5 ℃, exhibiting significantly better regulation stability than PID (Proportional-Integral-Derivative)and MPC (Model Predictive Control) strategies. The innovation contribution of this paper lies in the establishment of a complete prediction-analysis-regulation closed-loop chain. The adoption of SHAP interpretable analysis instead of traditional empirical rules effectively reduces the subjectivity of fuzzy control and improves the engineering adaptability and interpretability of regulation strategies.The innovation of this study is the establishment of a prediction-analysis-regulation. The proposed method can serve as a generalized scheme for the intelligent regulation of blast furnace molten iron temperature. Future research will combine reinforcement learning to realize the dynamic iteration of regulation rules and promote the extended application of the method in multi-blast furnace and cross-working-condition scenarios.