Deep learning has proven to be a valued contributor to recent technological advancements within energy systems. This thesis project explores methods of photovoltaic (PV) system power output forecasting through the utilization of long short-term memory (LSTM) neural networks. An encoder-decoder architecture (ED-LSTM) and a stacked vector output architecture (SVO-LSTM) were compared in terms of their ability to accurately produce power output forecasts with a 24-hour forecast horizon. The datasets which were used for model training were composed of historical meteorological observations and PV system power output readings. The results indicate that the encoder-decoder model and the stacked vector output model were somewhat equally skilled at producing power output forecasts. Best results were obtained by the encoder-decoder LSTM model which achieved a 26.63% improvement over a persistence model when trained on data sequences which preceded the forecast horizon, and a 44.96% improvement over a persistence model when the model was provided meteorological data from an oracle forecaster.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:uu-455665 |
| Date | January 2021 |
| Creators | Hamberg, Lukas |
| Publisher | Uppsala universitet, Institutionen för informationsteknologi |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | UPTEC IT, 1401-5749 ; 21008 |
Page generated in 0.0017 seconds