Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology
Thesis (PhD)--Stellenbosch University, 2021.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2021
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867614135500406784 |
|---|---|
| access_status_str | Open Access |
| author | Du Plessis, Armand |
| author2 | Rix, Arnold J. |
| author_browse | Du Plessis, Armand Rix, Arnold J. |
| author_facet | Rix, Arnold J. Du Plessis, Armand |
| author_sort | Du Plessis, Armand |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (PhD)--Stellenbosch University, 2021. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/109884 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA |
| last_indexed | 2026-06-10T12:47:13.687Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2021 |
| publishDateRange | 2021 |
| publishDateSort | 2021 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/109884 Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology Du Plessis, Armand Rix, Arnold J. Strauss, Johann M. Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Forecasting -- Statistical methods UCTD Photovoltaic power systems Power supply Energy consumption -- Forecasting Thesis (PhD)--Stellenbosch University, 2021. ENGLISH ABSTRACT: Power delivered from utility-scale Photovoltaic (PV) systems is characteristically intermittent, due to a dependence on atmospheric variables. To manage this uncertainty of an intermittent PV power supply, researchers traditionally adopt a macro-level forecasting approach, where a single model is trained to emulate the behaviour of the entire PV system. However, as commercial PV systems continue to expand in size, there is a growing uncertainty regarding the ability of these macro-level models to capture the non-uniform, low-level power output dynamics of large multi-megawatt PV systems. In response to this knowledge gap, a novel aggregated low-level forecasting methodology is proposed. With state-of-the-art deep learning (DL) implementations of Feedforward Neural Network (FFNN), Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) and Gated Recurrent Unit Recurrent Neural Network (GRU-RNN) models, the proposed methodology is compared to the conventional macro-level forecasting approach. With data obtained from a commercial 75 MW PV system, multi-step 1 - 6 h ahead forecasts are delivered for a realworld scenario. Forecast models are trained for each of the 84 inverters, which collectively serve as the aggregated low-level forecasting solution. However, given the high computational expense of training multiple forecast models, a unique and scalable inverter-clustering approach towards model development is presented. The discrepancies in literature concerning biased model development are also addressed, with a heuristic process of systematic hyperparameter optimisation proposed, which serves to guide future forecasting practitioners. Concerning the results, this research successfully demonstrates the application of the proposed methodology. From the day-time-only forecast results, the aggregated inverter-level FFNN model shows the largest improvement, with a Mean Absolute Percentage Error (MAPE) of between 0.04 % - 0.4 % lower in comparison to the FFNN macro-level forecasts. This translates to an overall 30 kW - 300 kW improvement in forecasting accuracy. The aggregated GRU-RNN inverter-level model forecasts deliver a smaller overall MAPE performance increase, ranging between 0.03 % - 0.1 %. This is a 20 kW - 75 kW improvement. However, compared to all the DL forecast models applied, the low-level GRU-RNN model forecasts deliver the highest overall forecasting accuracy, with MAPE values ranging between 5.8 % - 8 % (day-time-only forecasts). From the 95 % Bootstrap con dence intervals, no improvements regarding the uncertainty analysis are observed for the aggregated low-level forecasting methodology. Finally, with this research it is concluded that researchers who have and continue to propose DL-based forecast model solutions for smaller multi-megawatt PV systems, can be con dent in the application of these models as macro-level solutions. AFRIKAANSE OPSOMMING: Die konvensionele benadering om voorspellingsmodelle te ontwikkel, wat die drywingsuitset van groot multi-megawatt Fotovoltaïese (FV) kragstelsels voorspel, word baseer op 'n makrovlakvoorspellingsmetodologie, waar 'n enkele model geleer word om die gedrag van die hele FV-stelsel te voorspel. Soos wat kommersiële FV-stelsels uitbrei in grootte is daar egter 'n toenemende onsekerheid rakende die vermoë van hierdie makrovlakmodelle om die nie-uniforme, lae-vlak drywingsuittreedinamika vas te vang van hierdie groot FV-stelsels. In antwoord op hierdie kennisgaping word 'n nuwe FV-voorspellingsmetodologie voorgestel, waar dit ondersoek word of 'n verbeterde voorspellingsakkuraatheid bereik kan word met 'n samevoeging van veelvuldige lae-vlak voorspellingsmodelle. Dit is belangrik, omdat verdere verbetering wat in voorspellingsakkuraatheid gemaak word krities is vir elektrisiteitsnetwerkoperateurs, wie die toegevoegde druk van 'n wisselvallige FV-stelsel as energiebron e ektief moet bestuur. Om die fokus van die navorsing op die voorgestelde metodologie te behou, word geen hibriede modelle oorweeg nie, met slegs alleenstaande, mees gevorderde Diep-Leer (DL) modelle, wat insluit 'n Voer-Vorentoe-Neurale-Netwerke (VVNN), Lang-Korttermyngeheue-HerhalendeNeurale-Netwerke (LKTGHNN) en Hek-Herhalende-Eenheid-Herhalende-Neurale-Netwerke (HHEHNN), wat toegepas word. Met hierdie DL-modelle word multistap voorspellings 1 - 6 h vooruit gelewer vir 'n 75 MW netwerkgeskakelde FV-stelsel. Voorspellingsmodelle is vir elkeen van die 84 wisselrigters van die FV-stelsel ontwikkel, wat sodoende dien as die saamgevoegde laevlak voorspellingsoplossing. Die berekeningsuitdaging aangaande die skaleerbaarheid en reproduseerbaarheid van die verskeie DL-gebasseerde modelle is ook suksesvol aangespreek met 'n unieke wisselrigter-groeperingstegniek. Verder, as antwoord op die geïdenti seerde ongelykheid in navorsing in die literatuur, rakende die onregverdige en onoortuigende aanspraak van model meerderwaardigheid, word 'n heuristiese proses van stelselmatige hiperparameter-optimalisering voorgestel, wat dien as riglyn tot onbevooroordeelde model ontwikkeling vir die FV-voorspellingspraktisyn. Hierdie navorsing demonstreer die toepassing van die voorgestelde metodologie suksesvol. Die resultate van die daglig-alleen-voorspellings wys dat die saamgevoegde wisselrigtervlakVVNN-model die grootste verbetering toon, met 'n Gemiddelde-Absolute-Persentasie-Fout (GAPF) van tussen 0.04 % - 0.4 % laer in vergelyking met makrovlak VVNN-voorspellingsmodel. Dit kom neer op 'n algehele 30 kW - 300 kW verbetering in voorspellingsakkuraatheid. Die saamgevoegde wisselrigtervlak-HHEHNN-model lewer egter 'n kleiner algehele GAPFprestasie-verbetering, wat wissel tussen 0.03 % - 0.1 %. Dit is 'n 20 kW - 75 kW verbetering. Nietemin, in vergelyking met al die toegepaste DL-voorspellingsmodelle, het die saamgevoegde wisselrigtervlak-HHEHNN-model die hoogste algehele voorspellingsakkuraatheid bereik, met GAPF waardes wat wissel tussen 5.8 % - 8 % (daglig-alleen-voorspellings). Met die 95 % Bootstrap-sekerheidsinterval is daar geen verbeteringe aangaande die onsekerheidsanalise gevind vir die saamgevoegde laevlak-voorspellingsmetodologie nie. Laastens, met hierdie navorsing is daar tot die gevolgtrekking gekom dat navorsers, wie alreeds in die verlede, of nog beplan om nuwe DL-modelle te ondersoek as voorspellingsoplossings vir kleiner multi-megawatt FV-stelsels, met selfvertroue die toepassing van hierdie modelle as makrovlak-oplossings kan aanwend. Doctoral 2021-03-01T07:15:49Z 2021-04-21T14:30:19Z 2021-03-01T07:15:49Z 2021-04-21T14:30:19Z 2021-03 Thesis http://hdl.handle.net/10019.1/109884 en_ZA Stellenbosch University 205 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Forecasting -- Statistical methods UCTD Photovoltaic power systems Power supply Energy consumption -- Forecasting Du Plessis, Armand Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology |
| title | Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology |
| title_full | Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology |
| title_fullStr | Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology |
| title_full_unstemmed | Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology |
| title_short | Short-term power output forecasting for large multi-megawatt photovoltaic systems with an aggregated low-level forecasting methodology |
| title_sort | short term power output forecasting for large multi megawatt photovoltaic systems with an aggregated low level forecasting methodology |
| topic | Forecasting -- Statistical methods UCTD Photovoltaic power systems Power supply Energy consumption -- Forecasting |
| url | http://hdl.handle.net/10019.1/109884 |
| work_keys_str_mv | AT duplessisarmand shorttermpoweroutputforecastingforlargemultimegawattphotovoltaicsystemswithanaggregatedlowlevelforecastingmethodology |