Previsão do tempo de parada para decantação de sedimentos e limpeza das grades de proteção das unidades geradoras em usinas hidrelétricas utilizando modelos ocultos de Markov e redes bayesianas

Abstract

Hydropower plants installed in the Amazon basin are negatively influenced by transporting logs and sediments from the river bed. All transported material accumulates in the protection grids, which are installed in the water intake of the generating units to prevent the entry of this material and avoid damage to the turbines, which would result in economic losses. The accumulation of material on the protection grids reduces the water drop and, consequently, the flow available for generation, which results in load losses, preventing the generation unit from operating at full capacity. The presented problem occurs at the Jirau hydropower plant, located on the Madeira River and has 50 generation units, each with the potential to generate 75MW at total capacity. This plant operates on a run-of-river basis, meaning there is no reservoir for water storage. Therefore, all water resources must be used when available. The fact that it operates on a run-of-river, together with the problem of transporting logs and sediments, brings challenges to the operation of the plant, as the sediments that accumulate in the protection grids and consequently reduce the generation potential require the complete shutdown of the generating units so that the accumulated dirt can decant, thus enabling the resumption of the waterfall, and consequently the flow available for generation. A large number of generation units and the different location dispositions of equipment in the river alter the sediment accumulation profile, making it difficult to define rules that define the necessary stoppage time for each unit to decant the dirt and restart generation at total capacity. In this scenario, instead of relying only on the experience of operators, developing efficient methods capable of determining the ideal stoppage time for each generation unit becomes essential. In this work, prediction models using Bayesian Networks and Hidden Markov Models are proposed to estimate the downtime required for decanting the dirt from each generation unit so that it can be used again for generation in the shortest possible time. Big Data and Analytics techniques are also used to collect and process the large volume of data existing at the hydroelectric plant. The results demonstrate that the developed models could satisfactorily infer the time required for sediment decantation. The resulting model makes it possible to query information using various information, including the obstruction level when a unit stops, the obstruction level at restart time, whether neighboring units are operating, and in which power range.