Industrial production often involves complex time-varying operating conditions that result in continuous time-series production data. The traditional soft sensor approach has difficulty adjusting to such dynamic changes, which makes model performance less optimal. Furthermore, online analytical systems have significant operational and maintenance costs and entail a substantial delay in measurement output, limiting their ability to provide real-time control. In order to deal with these challenges, this paper introduces a multivariate multi-step predictive multilayer perceptron regression soft-sensing model, referred to as incremental MVMS-MLP. This model incorporates incremental learning strategies to enhance its adaptability and accuracy in multivariate predictions. As part of the method, a pre-trained MVMS-MLP model is developed, which integrates multivariate multi-step prediction with MLP regression to handle temporal data. Through the use of incremental learning, an incremental MVMS-MLP model is constructed from this pre-trained model. The effectiveness of the proposed method is demonstrated by benchmark problems and real-world industrial case studies.