The paper presents Soft Relational Temporal Cognitive Models (SRTCM), which combine the advantages of fuzzy cognitive models and machine learning models. These models are designed to perform a set of problems of scenario modeling and predictive analytics of complex systems, processes and problem situations in conditions of inaccuracy, incompleteness, uncertainty of data and nonlinear interdependence between heterogeneous systemic and external factors. The following issues are considered: designing the structure of SRTCM; selecting models of scenario dynamics for SRTCM concepts; structural and parametric settings of SRTCM using training samples for each of the concepts; setting modeling scenarios; scenario modeling and predictive analytics based on trained SRTCM. An example of scenario modeling of electrical load distribution for consumers in the regional energy subsystem using the SRTCM is illustrated. The formulation of the inverse problem of scenario modeling is presented, which consists in determining all the retrospective values of the SRTCM concepts in the time range, leading to the required values of the concepts at the time of scenario modeling. Various approaches are considered that expand the capabilities of SRTCM for diagnosing the state of complex systems and problem situations, process planning, resource management, with a given “time depth” of retrospective diagnosis.