In complex industrial processes, significant shifts in operating conditions lead to data distribution variations that challenge the accuracy of a single steady-state prediction model. Moreover, as historical data from previous conditions are often inaccessible, it is difficult to build a reliable new model with only limited data from the new condition. To address these issues, a soft sensor model reuse method based on reduced kernel mean embedding and broad learning system is proposed. The method operates in two stages: training and application. In the training stage, a broad learning system is used to develop a soft sensor model for historical operating conditions. A reduced set based on kernel mean embedding is designed to map historical data into a reproducing kernel Hilbert space. This preserves data privacy while capturing distribution characteristics. The model and its corresponding reduced set together form a model library. In the application stage, a distance metric based on maximum mean discrepancy is constructed to identify the most suitable model from the library. A distance threshold is then designed to dynamically update the model parameters for new operating conditions. The effectiveness and superiority of the proposed method are validated through two industrial case studies: a sulfur recovery process and a steelmaking process.