Developing hybrid deep learning models with data fusion approach for electricity consumption forecasting

Date

2023-9-06

Author

Özen, Serkan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

174
views

98
downloads

Cite This

Many of the deep learning solutions for electricity consumption forecasting reported in the literature include complex neural networks that may not be directly employed by the practitioner in the field. Throughout this study, we first demonstrate how the standard deep neural networks, i.e. convolutional neural network (CNN), long short-term memory (LSTM), and primitive methods, i.e. arima, random forest perform on univariate electricity consumption dataset. Then, we build hybrid models in order to test them on the newly formed multivariate dataset by combining weather and electricity datasets and show that they perform better on this dataset than the single models do on the univariate dataset. After doing this, we propose a hybrid model that utilizes data fusion approach, called shortly Fusion model. Fusion model consists of a single model that runs on a univariate dataset, a hybrid model that runs on a multivariate dataset and a linear regression model which is fed with the outputs of the single and hybrid models. As a result, we show that the overall result of Fusion model is better than any submodels’ results. The proposed Fusion model improves the average RMSE score to 0.0732 when compared to CNN, CNN+LSTM, LSTM+LSTM, kCNN-LSTM and naive Transformer models on Pittsburgh, Chicago and IHEC datasets. Moreover, we show the usability of transfer learning in case of lack of data, size of which is not necessary to fully train a model. Lastly, we attach our findings related to effect of preprocessing techniques and hybridization of transformer model for the electric consumption forecasting task.

Subject Keywords

time series forecasting, deep learning, hybrid models, data fusion

URI

https://hdl.handle.net/11511/105382

Collections

Graduate School of Natural and Applied Sciences, Thesis