Uma análise comparativa entre diferentes paradigmas de construção de modelos preditivos para vazão de reservatórios

Abstract

This study investigates one-step-ahead forecasting of mean monthly natural inflow in Brazilian hydroelectric reservoirs, an activity directly related to the planning and operation of the National Interconnected System (SIN). A comparative analysis is conducted among three data-driven approaches: a regularized autoregressive model, a recurrent neural network based on long and short-term memory, long short-term memory (LSTM), and a prompt-based approach using a large language model (LLM), specifically Gemini 2.5 Flash (GEMINI-2.5-F) and Gemini 3 Flash Preview (GEMINI-3-FP). A reproducible experimental protocol is adopted, employing sliding-window validation and chronological data splitting, applied to four reservoirs with distinct hydrological regimes: Furnas, Sobradinho, Itaipu, and Ilha Solteira. Model performance is evaluated on the original scale using the root mean square error (RMSE) and the coefficient of determination (R²). The results indicate that no single method achieves superior performance across all analyzed reservoirs. The LSTM yields lower forecasting errors in Furnas and Sobradinho, whereas GEMINI-3-FP attains better results in Itaipu and Ilha Solteira. These findings indicate that LLM-based approaches, even without direct parametric training on numerical time series, can produce consistent forecasts when the problem is formulated as inference conditioned on structured textual contexts. The study highlights the relevance of treating LLMs as a complementary approach to traditional hydrological modeling, whose application depends on the characteristics of the time series and requires careful empirical validation.