Data driven application for hydro generation
Tutor / Supervisor
Student
Domingo Torremorell, Josep
Document type
Master thesis
Date
2024
rights
Open Access
Publisher
Universitat Politècnica de Catalunya
UPCommons
Abstract
This thesis explores the effectiveness of various forecasting models for predicting the flow of a specific location of the river Flamicell, a critical component in hydroelectric power management. A comprehensive analysis was conducted on multiple models, including Linear and Ridge Regression, Random Forest, ARIMA, Auto-ARIMA, SARIMAX, Support Vector Regression, and neural network models such as RNN, LSTM and FNN. The study aimed to identify the most suitable model for accurately forecasting water flow in a hydroelectric context, given the complex and non-linear nature of the data. The results revealed that the Support Vector Regression model outperformed all other models, demonstrating the lowest error metrics and the highest R² Score, indicating its superior ability to capture the patterns and dynamics inherent in the Flow Flamicell data. The Random Forest model also showed strong performance, particularly in short-term forecasting, though it did not match the precision of the SVR model. Traditional statistical models such as ARIMA, Auto-ARIMA, and SARIMAX were found to be less effective due to their limited ability to generalize to unseen data and handle non-stationary and complex temporal dependencies. Neural network models, particularly the RNN, also showed promise in capturing sequential dependencies, though their performance did not exceed that of the SVR model.
