Assessing power system reliability under temperature-dependent variables – A case study of the Brazilian Power System
This work focuses on explicitly representing key variables of the power sector under temperature-dependent conditions, with particular emphasis on electricity demand and generator availability. By jointly modeling these two critical dimensions, the thesis seeks to evaluate the reliability of the Brazilian National Interconnected System (SIN) under stress conditions, particularly in the context of growing exposure to climatic extremes.
The SIN, coordinated by the Independent System Operator (ONS), supplies nearly the entire country, except for the state of Roraima, which is still electrically isolated but expected to be interconnected by the end of 2025. The system is divided into four major subsystems (SE/CO, S, NE, N), each with distinct socioeconomic, climatic, and load characteristics. The relevance of demand forecasting and generator availability analysis lies not only in their role for short-term operations, but also in medium- and long-term planning, capacity expansion, and market mechanisms. Currently, the country is discussing the implementation of capacity auctions, which will seek to contract firm capacity, mostly from thermal plants, to meet demand at critical times. Incorporating temperature into these variables is particularly critical in Brazil, where heat waves and seasonal variability significantly affect both consumption patterns and thermal unit performance.
To address this challenge, a comprehensive and historically consistent database was compiled. For demand, semi-hourly records from the ONS API were compiled and complemented by the identification of critical daily peaks. For climate, hourly records of temperature from the National Institute of Meteorology (INMET) were collected, consolidated, and treated with statistical methods, seeking to solve the missing data set and apply a clustering process to select representative weather stations for each subsystem, ensuring both spatial and climatic diversity.
On the demand modeling side, two families of methods were applied. Econometric regression (OLS) and time-series models (ARIMAX and SARIMAX) were used to forecast monthly demand and critical daily profiles, incorporating economic indicators, temperature-based variables such as Heating and Cooling Degree Days, and seasonal harmonics. Variable selection was performed using backward elimination and LASSO regularization to enhance interpretability and robustness. Finally, Monthly dummies variables and an alternative with Dynamic Harmonic Regression was applied to represent the seasonality. A stochastic scenario generation procedure was implemented through a double bootstrap of historical climate data, preserving temporal
correlations across stations and allowing for the inclusion of linear warming trends, in order to provide future temperature scenarios and in this sense analyze the probabilistic distribution of the forecasted demand.
The results demonstrate that explicitly incorporating temperature improves predictive accuracy of demand by reducing RMSE by over 38% on average in the monthly forecast demand process. Also, superior results are obtained when ARIMA models are applied, compared with traditional OLS and with the SARIMA models. In the hourly forecast assessment, the impacts by considering the temperature are clear, with some months presenting differences above 10% for the same hour under different temperature conditions.
Finally, the reliability of the system is assessed under the generated forecast demand for 2028, considering a temperature-dependent outage probability for the thermal generators. For this purpose, taking as a starting point the database made available from the official forecasts from the Brazilian System Operator, the capacity requirement for the thermal fleet is estimated, and subsequently stress tested considering the temperature scenarios. The methodology applies a cut-based approach to the Brazilian interconnected system, accounting for hydroelectric reservoirs with explicit regulation constraints, stochastic renewable generation scenarios built from ERA5 reanalysis data through the TSL platform, and a gamma distribution to represent the aggregated expected forced outage of thermal capacity per subsystem. In total, one hundred inflow and renewable scenarios are combined with one hundred peak demand scenarios, focusing on the critical hours of supply between 5 p.m. and 2 a.m., generating a broad set of stress conditions.
The results confirm that the most critical deficits occur at sunset (5–7 p.m.), when solar falls while demand remains high. Deficits exceed 10 GW and surpass 13 GW in extreme scenarios, while even median cases reveal structural deficits above 5 GW. When thermal failures are modeled as temperature dependent, the capacity requirement rises by about 500 MW, highlighting the systemic risks of ignoring climate dependency. Economically, this additional capacity equates to R$ 412 million per year under current firm capacity benchmarks, rising to nearly R$ 2 billion if contracted under emergency auctions, as occurred in 2021. These findings stress the need to explicitly integrate climate dependencies in planning and underscore the importance of adding flexible capacity capable of supporting the system during sunset peaks.
