Stochastic Methods for Forecasting and Power System Mode Optimization with a High Share of Renewable Energy Sources: A Case Study of Tajikistan

Subject. This paper addresses the critical scientific challenge of integrating stochastic solar generation into the power system of Tajikistan, which is predominantly characterized by hydroelectric power. The high volatility of solar irradiance in complex terrain poses risks to static stability and necessitates a transition from deterministic operational planning methods. Methods. To solve the short-term forecasting problem, the study proposes an original hybrid approach based on the synthesis of Long Short-Term Memory (LSTM) recurrent neural networks and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) econometric models. The theoretical framework relies on the decomposition of power time series and entropy analysis of the information complexity of insolation patterns. Results. An adaptive two-loop forecasting algorithm was developed, enabling both the approximation of non-linear production trends and the generation of dynamic confidence intervals based on current volatility estimates. Testing the model on empirical data for 2021–2023 demonstrated a reduction in the Mean Absolute Percentage Error (MAPE) to 6.4%. This outperforms classical ARIMA methods and standalone neural network models by 15–20% during periods of unstable cloud cover. Practical Significance. The findings allow for the optimization of spinning reserves at cascading hydropower plants and improve the efficiency of water-energy resource management during seasonal power shortages. The proposed architecture is ready for integration into modern Automated Dispatch Control Systems (ADCS) via standard data exchange protocols.