IMPLEMENTATION OF MULTIDIMENSIONAL SYSTEMS IN A RELAXATION ENVIRONMENT BASED ON LSTM MODELING FOR MONITORING AND FORECASTING WATER RESOURCES

Abstract

This paper examines the automated control of closed reservoir systems (CRS) (using endorheic lakes as an example) operating under stochastic hydrological dynamics and characterized by internal relaxation processes (non-Markovian behavior). An intelligent stochastic control system is proposed based on relaxation-stochastic models (RSMs), which take into account both the probabilistic nature of external influences and the system's internal memory. The risk-based optimal control problem is solved in real time using recessing horizon control (RHC) and an LSTM-based metamodel approximating the complex dynamics of RCMs. Testing the method on the Aydarkul Lake system (Uzbekistan) demonstrated a 37% reduction in the probability of water levels falling below the critical ecological limit over a 10-year forecast period, while simultaneously increasing the reliability of irrigation water supply by 18% compared to traditional strategies. The obtained results demonstrate the effectiveness of the proposed approach for creating robust and adaptive water resources automation systems under conditions of high uncertainty.