Development of CNN-Based Models for Short-Term Load Forecasting in Energy Systems
| dc.contributor.advisor | Tkaczyk, Alan Henry | |
| dc.contributor.advisor | Dam, Florentin | |
| dc.contributor.advisor | Aakenes, Ulf Roar | |
| dc.contributor.author | Kurylenko, Oleksandr | |
| dc.contributor.other | Tartu Ülikool. Loodus- ja täppisteaduste valdkond | et |
| dc.contributor.other | Tartu Ülikool. Tehnoloogiainstituut | et |
| dc.date.accessioned | 2021-05-31T07:33:33Z | |
| dc.date.available | 2021-05-31T07:33:33Z | |
| dc.date.issued | 2020 | |
| dc.description.abstract | In this work, two deep learning models based on convolutional neural networks (CNNs) are developed for one-day-ahead hourly electricity consumption forecasting for the next calendar day. One-day-ahead electricity consumption forecasting is done to perform baseline load evaluation for the assessment of Demand Response (DR) performance and to predict the availability of flexibility products. Using electricity consumption time series data for three regions in Norway, the developed CNNbased models are compared to four industry-standard baseline models (Asymmetric High Five of Ten, Similar Profile Five of Ten, Average, and Daily Profile). Additionally, comparisons are made to a Naive model and an Autoregressive Integrated Moving Average (ARIMA) model, which includes Fourier terms. Three evaluation metrics are used to estimate the models’ performance. A detailed description of the methodology, work pipeline, and results is provided. The conducted experiments were successful in developing and applying the CNN-based models to the problem of one-day-ahead hourly electricity consumption forecasting for the next calendar day. The developed CNN and combination “CNN + Long Short-Term Memory (LSTM)” models showed the best performance results among the predictive models which employ the Multiple-Input Multiple-Output (MIMO) strategy to forecast 24 hours ahead. The Daily Profile model showed the best performance among models which cannot forecast 24 hours ahead or do not necessarily employ the MIMO strategy. It is worth noting that the Average model showed the best performance in the baseline load evaluation among all the considered models. However, the Average model is only a reference comparison which does not actually perform forecasting, but rather uses post-event data. It can be concluded that the work was successful in developing and applying the CNN-based models to short-term load forecasting in energy systems, especially since the developed CNN and CNN+LSTM models outperformed other similar forecasting models. Two different paths could be chosen for future work: one that intends to explore more and improve the CNN-based models developed in this work, and the one which aims to explore new CNN-based architectures. | en |
| dc.identifier.uri | http://hdl.handle.net/10062/72116 | |
| dc.language.iso | eng | et |
| dc.publisher | Tartu Ülikool | et |
| dc.rights | embargoedAccess | et |
| dc.rights.uri | http://creativecommons.org/licenses/by-nc-nd/3.0/ee/ | |
| dc.subject | konvolutsioonilised närvivõrgud (CNN) | et |
| dc.subject | pikk lühiajaline mälu (LSTM) | et |
| dc.subject | ühe muutujaga aegrea prognoosimine | et |
| dc.subject | mitme muutujaga aegrea prognoosimine | et |
| dc.subject | elektri tarbimine | et |
| dc.subject | elektri nõudlus | et |
| dc.subject | Convolutional Neural Network (CNN) | en |
| dc.subject | Long Short-Term Memory (LSTM) | en |
| dc.subject | univariate time series forecasting | en |
| dc.subject | multivariate time series forecasting | en |
| dc.subject | electricity consumption | en |
| dc.subject | electricity demand | en |
| dc.subject.other | magistritööd | et |
| dc.title | Development of CNN-Based Models for Short-Term Load Forecasting in Energy Systems | en |
| dc.title.alternative | Konvolutsioonilistel närvivõrkudel põhinevate mudelite arendus lühiajalise koormuse prognoosimiseks energiasüsteemides | et |
| dc.type | info:eu-repo/semantics/masterThesis | et |