Attention-based multi-task deep learning for short-term multi-energy load forecasting
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Insinööritieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2024-06-10
Department
Major/Subject
Mcode
Degree programme
Environomical Pathways for Sustainable Energy Systems
Language
en
Pages
72+5
Series
Abstract
Finland produces 38.1 TWh of district heat annually, making it the leading producer of district heat per capita among the Nordic countries. Additionally, it generates 82 TWh of electricity per year, with 41.8% coming from renewable sources, placing it fifth among the Nordic countries. The country's significant role in district heating and its potential to increase renewable energy integration underscores the importance of an energy transition for optimal energy use and sustainable development. Although the energy transition in Finland is complex and influenced by multiple factors, effective decision-making is essential for facilitating this transformation. A key aspect of such decision-making occurs in the domain of day-ahead optimal dispatch within Multi-Energy Systems (MES). This process involves making forward-looking decisions on aspects such as electricity pricing, renewable energy integration, and operations for the following day. Accurate multi-energy short-term load forecasting is crucial in this context, as it provides the necessary information for optimal decision-making, thereby improv-ing efficiency, economic outcomes, and sustainability within MES. Therefore, this thesis seeks to establish an accurate Multi-energy Load Forecasting (MELF) framework using a comprehensive, machine learning-based approach. It validates the effectiveness of multi-task learning (MTL) in predicting multi-energy loads. Through three chapters of step-by-step validation, the study examines 1) traditional machine learning methods, 2) single task deep learning methods, and 3) MTL methods. The results demonstrate that MTL methods can effectively predict multi-energy loads, with Bidirectional GRU (BiGRU) as the best-performing model. An attention mechanism was employed within the MTL framework to avoid overfitting, addressing the challenge of managing a large number of features in multi-task learning objectives. Finnish community load data was used to validate the proposed frame-work. The MTL-BiGRU model significantly outperforms other models, achieving the lowest Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE) for electricity (0.5882%, 4.8618) and heat load (2.4660%, 11.3059) predictions, along with the highest R² values (0.9980, 0.9977). These results demonstrate superior accuracy, substantial error reduction, and excellent model fit.Description
Supervisor
Li, ZhengmaoThesis advisor
Jia, XueyongKeywords
multi energy load forecasting, multi task learning, multi energy system, attention mechanism, machine learning, deep learning