Monthly Runoff Forecasting by Non-Generalizing Machine Learning Model and Feature Space Transformation (Vakhsh River Case Study)

Abstract

Energy prices and сost of materials for solar and wind power plants have increased over the past year. Therefore, significance increases for the hydropower and long-term (1–10 years) planning generation for the existing hydropower plants, which requires forecasting the average monthly values of the river flow. This task is especially urgent for countries without their own oil-fields and opportunity to invest in the creation of solar or wind power plants. The aim of the research is to decrease the mean absolute forecasting error of the long-term prediction for the Vakhsh River flow (Tajikistan) based on the long-term observations. A study of existing methods for the river runoff forecasting in relation to the object under consideration was carried out, and a new transformation model for the space of the input features was developed. The most significant results are the decrease in the average forecast error in the Vakhsh river flow achieved by the use of the proposed space of polynomial logarithmic features in comparison with other methods, and the need to use at least the 20 year-old observational data for the long-term operation planning of the hydropower plants and cascades of the hydropower plants obtained from the results of computational experiments. The significance of the results lies in the fact that a new approach to the long-term forecasting of river flow has been proposed and verified using the long-term observations. This approach does not require the use of the long-term meteorological forecasts, which are not possible to obtain with high accuracy for all regions.