Deakin microgrid digital twin and analysis of AI models for power generation prediction
To achieve carbon neutral by 2025, Deakin University launched a AUD 23 million Renewable Energy Microgrid in 2020 with a 7-megawatt solar farm, the largest at an Australian University. A web-based digital twin (DT) is developed to provide operators with intelligence and insights through several AI-d...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2023-04-01
|
| Series: | Energy Conversion and Management: X |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2590174523000260 |
| _version_ | 1797842750628954112 |
|---|---|
| author | Iynkaran Natgunanathan Vicky Mak-Hau Sutharshan Rajasegarar Adnan Anwar |
| author_facet | Iynkaran Natgunanathan Vicky Mak-Hau Sutharshan Rajasegarar Adnan Anwar |
| author_sort | Iynkaran Natgunanathan |
| collection | DOAJ |
| description | To achieve carbon neutral by 2025, Deakin University launched a AUD 23 million Renewable Energy Microgrid in 2020 with a 7-megawatt solar farm, the largest at an Australian University. A web-based digital twin (DT) is developed to provide operators with intelligence and insights through several AI-driven capabilities. Accurate and computationally efficient power generation prediction is one of the critical elements in this DT. To this end, we researched the literature and identified the commonly used Machine Learning-based prediction models and compared them computationally using power generation and weather sensor data obtained from the solar farm. From the computational experiments, we find that, overall, Artificial Neural Network (ANN) has achieved the highest R2-score (0.944) and the lowest RMSE (14.848). To obtain further insights, we compared the methods using our two novel metrics, the x-percentile Closeness scores and the x-percentile Absolute error scores. The new metrics provide us with a spectrum to measure the consistency and robustness of the prediction methods instead of just a single value. Further, power generation can fluctuate substantially, and a prediction model should be accurate regardless of the magnitude of the output, hence measuring the relative error has its merits. By our two new metrics, using the data from our Deakin Microgrid, Random Forrest (RF) outperformed the other methods tested, with the smallest absolute relative error across the whole spectrum (from 0.011 to 0.457). RF is also the fastest in model training time at 4.894 s and XGBoost came second at 5.115 s–a big contrast to ANN at 144.102 s. All prediction times are under 1 s. RF is therefore used as a power prediction algorithm in our Deakin Microgrid Digital Twin. |
| first_indexed | 2024-04-09T16:53:06Z |
| format | Article |
| id | doaj.art-099693775dcb4019944a250c82378898 |
| institution | Directory Open Access Journal |
| issn | 2590-1745 |
| language | English |
| last_indexed | 2024-04-09T16:53:06Z |
| publishDate | 2023-04-01 |
| publisher | Elsevier |
| record_format | Article |
| series | Energy Conversion and Management: X |
| spelling | doaj.art-099693775dcb4019944a250c823788982023-04-21T06:46:12ZengElsevierEnergy Conversion and Management: X2590-17452023-04-0118100370Deakin microgrid digital twin and analysis of AI models for power generation predictionIynkaran Natgunanathan0Vicky Mak-Hau1Sutharshan Rajasegarar2Adnan Anwar3Deakin University, Geelong, VIC 3215, AustraliaCorresponding author.; Deakin University, Geelong, VIC 3215, AustraliaDeakin University, Geelong, VIC 3215, AustraliaDeakin University, Geelong, VIC 3215, AustraliaTo achieve carbon neutral by 2025, Deakin University launched a AUD 23 million Renewable Energy Microgrid in 2020 with a 7-megawatt solar farm, the largest at an Australian University. A web-based digital twin (DT) is developed to provide operators with intelligence and insights through several AI-driven capabilities. Accurate and computationally efficient power generation prediction is one of the critical elements in this DT. To this end, we researched the literature and identified the commonly used Machine Learning-based prediction models and compared them computationally using power generation and weather sensor data obtained from the solar farm. From the computational experiments, we find that, overall, Artificial Neural Network (ANN) has achieved the highest R2-score (0.944) and the lowest RMSE (14.848). To obtain further insights, we compared the methods using our two novel metrics, the x-percentile Closeness scores and the x-percentile Absolute error scores. The new metrics provide us with a spectrum to measure the consistency and robustness of the prediction methods instead of just a single value. Further, power generation can fluctuate substantially, and a prediction model should be accurate regardless of the magnitude of the output, hence measuring the relative error has its merits. By our two new metrics, using the data from our Deakin Microgrid, Random Forrest (RF) outperformed the other methods tested, with the smallest absolute relative error across the whole spectrum (from 0.011 to 0.457). RF is also the fastest in model training time at 4.894 s and XGBoost came second at 5.115 s–a big contrast to ANN at 144.102 s. All prediction times are under 1 s. RF is therefore used as a power prediction algorithm in our Deakin Microgrid Digital Twin.http://www.sciencedirect.com/science/article/pii/S2590174523000260Solar energyMicrogridPower predictionMachine learningDeep learning |
| spellingShingle | Iynkaran Natgunanathan Vicky Mak-Hau Sutharshan Rajasegarar Adnan Anwar Deakin microgrid digital twin and analysis of AI models for power generation prediction Energy Conversion and Management: X Solar energy Microgrid Power prediction Machine learning Deep learning |
| title | Deakin microgrid digital twin and analysis of AI models for power generation prediction |
| title_full | Deakin microgrid digital twin and analysis of AI models for power generation prediction |
| title_fullStr | Deakin microgrid digital twin and analysis of AI models for power generation prediction |
| title_full_unstemmed | Deakin microgrid digital twin and analysis of AI models for power generation prediction |
| title_short | Deakin microgrid digital twin and analysis of AI models for power generation prediction |
| title_sort | deakin microgrid digital twin and analysis of ai models for power generation prediction |
| topic | Solar energy Microgrid Power prediction Machine learning Deep learning |
| url | http://www.sciencedirect.com/science/article/pii/S2590174523000260 |
| work_keys_str_mv | AT iynkarannatgunanathan deakinmicrogriddigitaltwinandanalysisofaimodelsforpowergenerationprediction AT vickymakhau deakinmicrogriddigitaltwinandanalysisofaimodelsforpowergenerationprediction AT sutharshanrajasegarar deakinmicrogriddigitaltwinandanalysisofaimodelsforpowergenerationprediction AT adnananwar deakinmicrogriddigitaltwinandanalysisofaimodelsforpowergenerationprediction |