Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Time series forecasting using neural networks without recurrent connections
Masunungure, T. T. 2025. Time series forecasting using neural networks without recurrent connections. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/4fe53217-4ff1-4a6c-9b37-aa467356b164
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613876511571968 |
|---|---|
| access_status_str | Open Access |
| author | Masunungure, Takudzwa Trevor |
| author2 | Engelbrecht, Andries |
| author_browse | Engelbrecht, Andries Masunungure, Takudzwa Trevor |
| author_facet | Engelbrecht, Andries Masunungure, Takudzwa Trevor |
| author_sort | Masunungure, Takudzwa Trevor |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Masunungure, T. T. 2025. Time series forecasting using neural networks without recurrent connections. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/4fe53217-4ff1-4a6c-9b37-aa467356b164 |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/132630 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:43:06.958Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/132630 Time series forecasting using neural networks without recurrent connections Masunungure, Takudzwa Trevor Engelbrecht, Andries Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering. Time-series analysis -- Forecasting Prediction theory -- Mathematical models Deep learning (Machine learning) Neural networks (Computer science) UCTD Masunungure, T. T. 2025. Time series forecasting using neural networks without recurrent connections. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/4fe53217-4ff1-4a6c-9b37-aa467356b164 Thesis (MEng)--Stellenbosch University, 2025. ENGLISH ABSTRACT: Recent research has demonstrated that recurrent connections are not necessary in neural networks (NNs) for forecasting various time series. The approach defines the training of a NN for non-stationary time series prediction as a dynamic optimization problem (DOP) and proposes that training a NN using a quantum-inspired particle swarm optimization (QPSO) is sufficient for non-stationary time series forecasting. This study aimed to further explore the hypothesis that recurrent connections are not necessary by exploring the forecasting performance of QPSO trained NNs against that of NNs trained using standard particle swarm optimization (PSO), adaptive moment estimation (Adam), and resilient propagation (RPROP). Experiments were done on six forecasting problems. For each problem, nine NN models were trained using QPSO, standard PSO, Adam, and RPROP. The NN models used are feedforward NNs (FNNs), Elman NNs (ENNs), Jordan NNs (JNNs), multi-recurrent NNs (MRNNs), time delay NNs (TDNNs), long short-term memory (LSTM) networks, gated recurrent unit (GRU) networks, recurrent convolutional NNs (RCNNs), and temporal convolutional NNs (TCNNs). Each model was evaluated under three dynamic environments (DEs) illustrating different change severities and change frequencies. Results indicate that QPSO-trained FNNs outperform shallow recurrent NNs (SRNNs) trained using standard PSO, RPROP or QPSO. Additionally, the findings of the study showed that deep NNs (DNNs) trained using QPSO perform as well or better than DNNs trained using Adam. AFRIKAANSE OPSOMMING: Onlangse navorsing het getoon dat herhalende verbindings nie noodsaaklik is in neurale netwerke (NNe) vir die voorspelling van verskeie tydreekse nie. Die benadering definieer die opleiding van ’n NN vir die voorspelling van nie-stasionêre tydreekse as ’n dinamiese optimaliseringsprobleem (DOP) en stel voor dat die opleiding van ’n NN met behulp van ’n kwantum-geïnspireerde deeltjieswerm optimalisering (QPSO) voldoende is vir die voorspelling van nie-stasionêre tydreekse. Hierdie studie het daarop gemik om die hipotese verder te ondersoek dat herhalende verbindings nie noodsaaklik is nie, deur die voorspellingsprestasie van NNe wat met QPSO opgelei is, te vergelyk met dié van NNe wat met standaard deeltjieswerm-optimalisering (PSO), aanpasbare momentberaming (Adam), en veerkragtige propagering (RPROP) opgelei is. Eksperimente is op ses voorspellingsprobleme uitgevoer. Vir elke probleem is nege NN-modelle opgelei deur gebruik te maak van QPSO, standaard PSO, Adam, en RPROP. Die NN-modelle sluit in vorentoe-voer NNe (FNNe), Elman NNe (ENNe), Jordan NNe (JNNe), multi-herhalende NNe (MRNNe), tydsvertraging NNe (TDNNe), langkorttermyngeheue (LSTM) netwerke, omheinde herhalende eenheid (GRU) netwerke, herhalende konvolusie NNe (RCNNe), en temporale konvolusie NNe (TCNNe). Elke model is geëvalueer onder drie dinamiese omgewings (DOs) wat verskillende veranderingsintensiteite en veranderingsfrekwensies illustreer. Resultate toon dat FNNe wat met QPSO opgelei is, beter presteer as vlak herhalende NNe (SRNNe) wat met standaard PSO, RPROP of QPSO opgelei is. Verder het die bevindinge van die studie aangedui dat diep NNe (DNNe) wat met QPSO opgelei is, so goed of beter presteer as DNNe wat met Adam opgelei is. Masters 2025-06-12T06:55:39Z 2025-06-12T06:55:39Z 2025-03 Thesis https://scholar.sun.ac.za/handle/10019.1/132630 en Stellenbosch University xiv, 165 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Time-series analysis -- Forecasting Prediction theory -- Mathematical models Deep learning (Machine learning) Neural networks (Computer science) UCTD Masunungure, Takudzwa Trevor Time series forecasting using neural networks without recurrent connections |
| title | Time series forecasting using neural networks without recurrent connections |
| title_full | Time series forecasting using neural networks without recurrent connections |
| title_fullStr | Time series forecasting using neural networks without recurrent connections |
| title_full_unstemmed | Time series forecasting using neural networks without recurrent connections |
| title_short | Time series forecasting using neural networks without recurrent connections |
| title_sort | time series forecasting using neural networks without recurrent connections |
| topic | Time-series analysis -- Forecasting Prediction theory -- Mathematical models Deep learning (Machine learning) Neural networks (Computer science) UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/132630 |
| work_keys_str_mv | AT masununguretakudzwatrevor timeseriesforecastingusingneuralnetworkswithoutrecurrentconnections |