This study aims to enhance the accuracy of short-term load forecasting for a non-household electricity consumer in Ukraine — including institutions such as schools, hospitals, and government facilities — under conditions of high data variability, by applying deep Long Short-Term Memory (LSTM) neural networks. The specific contribution of this research does not consist in the mere application of the LSTM architecture, but rather in demonstrating its ability to reliably forecast highly unstable load time series of a non-household consumer under real operating conditions in Ukraine, including emergency outages and strong variability. The practical significance of the research consists in the potential application of the proposed approach to improve the reliability and accuracy of electrical load management systems in power distribution networks, including microgrids, without a substantial increase in computational costs. The study employs methods of statistical time series analysis, multifactor feature selection, and modeling based on deep LSTM neural networks, alongside classical approaches such as Singular Spectrum Analysis (SSA) and Holt–Winters exponential smoothing. Experiments were conducted for forecasting horizons of 1 h and 24 h. Forecasting results on real-world data indicate that the LSTM architecture yields lower forecasting error values (MAE, RMSE, MAPE) compared to classical models. Specifically, for the 1-h forecast, the MAPE decreased on average by 75.5%, and for the 24-h forecast – by 70.8%. These findings support the feasibility and potential of further improving the LSTM model for solving load forecasting tasks in power systems.