This article explores the development of methods for predicting the risk potential of technological processes at a VVER-1000 nuclear power plant unit based on the refinement of instantaneous predictive models. Building on a previous study on selecting a predictive model for condenser cooling water heating, an adaptive approach to generating instantaneous models is proposed to improve accuracy by dynamically selecting the optimal number of associative vectors. The model structure, including the prehistory depth, is determined at the initial stage through correlation analysis and remains fixed, whereas the number of associative vectors is adapted to the unit's current operating mode using a cross-validation criterion. The classification of instantaneous models based on associative search within machine learning paradigms is considered: it is shown that they represent a specific variant of instance-based learning, characterized by a two-stage procedure: non-chronological selection of precedents and the construction of a parametric model based on them at each cycle. Experimental studies using real power unit operating data confirmed that the adaptive model reduces the root-mean-square error in predicting cooling water heating by 18% compared to the baseline model with fixed parameters. The results of implementing this approach for the information task of calculating a unit's technical and economic indices are presented.