The article considers approaches to multi-level data processing using the Practical Byzantine Fault Tolerance (PBFT) and Reliable, Replicated, and Fault-Tolerant (RAFT) consensus algorithms. Their application for organizing efficient processing and aggregation of transactional data in distributed systems is analyzed. The main focus is on the use of recurrent neural networks with long short-term memory (LSTM) for predicting transactions based on data processed by PBFT and RAFT. The results of experiments on improving the accuracy of predictions taking into account various consensus strategies are presented.