Novosibirsk State Technical University (NETI) has proposed optimizing the operation of hydroelectric power plants using predictive models based on machine learning methods. This will make it possible to use water more efficiently and minimize idle discharges during flood periods.
"Hydroelectric power plants play an important role in the energy system, they have high maneuverability, can quickly gain power and are an environmentally friendly energy source and one of the cheapest in terms of maintenance and cost of energy resources. Forecasting the natural flow of water to the site of the hydroelectric power plant is one of the main elements of planning the operation of the power plant. The inflow of water in the river varies within a year with a certain frequency. The inflow of water in the river varies within a year with a certain frequency. There are two distinct periods — the period of autumn and the period of high water. In addition, the inflow of water varies over a long-term interval (low-water, medium-water and high-water years). If the year turned out to be low-water, the resource will be limited and we must use it as economically as possible to meet the requirements for the provision of guaranteed power to the power grid. The accuracy of forecasting water inflow is important for the reliable and efficient operation of a power plant," said Sergey Mitrofanov, Associate Professor of the Department of Power Supply Systems at NSTU-NETI, Candidate of Technical Sciences.
To solve the problem of planning the operation of the HPP, a sample was made from meteorological information for 9 years with values of air temperature, pressure, precipitation, humidity, and the average daily values of water consumption were also taken into account. The scientists used this information to create a forecasting model.The scientists used this information to create a forecasting mode. After testing different machine learning methods, they came to the conclusion that models based on a random forest of decision trees and gradient boosting would be the most effective. They allow us to consider several additional factors and limitations and generally demonstrate higher accuracy (the highest accuracy in the model of a random forest of decision trees was 86.52%).
The models are software modules. A set of data is fed to the input, then the parameters are calculated according to a certain algorithm, and the calculated value is obtained at the output. the task of the models is to predict the inflow of water to the HPP site, and the more accurate the forecast, the more effectively the plant's operating mode will be planned. Thus, NSTU-NETI scientists have modeled the most difficult period of pre-water operation (weekly and decadal forecast) for the Novosibirsk HPP.
As Sergey Mitrofanov noted, the practical significance of the development lies in the ability to choose the best operating mode of the HPP, considering the weather conditions. An accurate forecast makes it possible to prepare for changes in the inflow and consumption regime, develop a strategy for filling the reservoir in advance, and minimize the volume of idle discharges.
The results of the study were published in the International Journal of Hydrogen Energy.