Hydroelectric power generation is a critical component of the global energy matrix, particularly in countries like Brazil, where it represents the majority of the energy supply. However, its strong dependence on river discharges, which are inherently uncertain due to climate variability, poses significant challenges. River discharges are linked to precipitation patterns, making the development of accurate probabilistic forecasting models crucial for improving operational planning in systems heavily reliant on this resource. Traditionally, statistical models have been used to represent river discharges in energy optimization. Yet, these models are increasingly unable to produce realistic scenarios due to structural shifts in climate behavior. Changes in precipitation patterns have altered discharge dynamics, which traditional approaches struggle to capture. Machine learning methods, while effective as universal predictors for time series, often focus solely on historical data, ignoring key external factors such as meteorological and climatic conditions. Furthermore, these methods typically lack a probabilistic framework, which is vital for representing the inherent variability of hydrological processes. The limited availability of historical discharge data further complicates the application of large-scale deep learning models to this domain. To address these challenges, we propose a framework based on a modified recurrent neural network architecture. This model generates parameterized probability distributions conditioned on projections from global circulation models, effectively accounting for the stochastic nature of river discharges. Additionally, the architecture incorporates enhancements to improve its generalization capabilities. We validate this framework within the Brazilian Interconnected System, using projections from the SEAS5-ECMWF system as conditional variables.

翻译：水力发电是全球能源结构中的关键组成部分，在巴西等国家尤其如此，其占据了能源供应的主体地位。然而，水电高度依赖于河流流量，而由于气候的多变性，河流流量本身具有不确定性，这带来了重大挑战。河流流量与降水模式密切相关，因此，为严重依赖该资源的系统改进运行规划，开发精确的概率预测模型至关重要。传统上，在能源优化中常使用统计模型来表征河流流量。然而，由于气候行为的结构性变化，这些模型越来越难以生成符合现实的情景。降水模式的变化改变了流量动态，这是传统方法难以捕捉的。机器学习方法作为时间序列的通用预测器虽然有效，但往往仅关注历史数据，忽略了气象和气候条件等关键外部因素。此外，这些方法通常缺乏概率框架，而该框架对于表征水文过程固有的变异性至关重要。历史流量数据的有限可用性进一步阻碍了大规模深度学习模型在该领域的应用。为应对这些挑战，我们提出一个基于改进的循环神经网络架构的框架。该模型生成以全球环流模型预测为条件的参数化概率分布，从而有效解释河流流量的随机性。此外，该架构还结合了增强功能以提高其泛化能力。我们在巴西互联电力系统中验证了该框架，使用SEAS5-ECMWF系统的预测作为条件变量。