A solar forecasting model using First Nations seasonal knowledge is as much as 26 per cent more accurate than using traditional prediction models, a group of researchers in Darwin have found.
The Charles Darwin University team used machine learning to analyse seasonal calendars to predict solar panel output in the future.
“Incorporating First Nations seasonal knowledge into solar power generation predictions can significantly enhance accuracy by aligning forecasts with natural cycles that have been observed and understood for thousands of years,” the paper, published in IEEE Explore, says.
“The enhanced prediction ability suggests that integrating various First Nations seasonal information can significantly refine forecasting models. Moreover, these results highlight the potential for incorporating diverse and culturally relevant data to improve the performance of predictive analytics in future energy applications.”
The results were between 14 per cent and slightly more than 26 per cent more accurate when using that local knowledge.
The model used Tiwi, Gulumoerrgin, Kunwinjku and Ngurrungurrudjba First Nations calendars, a modern calendar called the Red Centre, and created a new dataset dubbed AliDKA using information from the Desert Knowledge Australia Solar Center in Alice Springs.
The FNS-Metrics (First Nations seasonal) captured temperature, irradiance, rainfall, and details of transitional weather patterns, while AliDKA covered temperature, relative humidity, two readings for horizontal radiation, wind direction, daily rainfall, and both global and diffuse tilted radiation
The team found they could achieve an error rate that is less than half of that in popular forecasting models used in the industry today.
“The … framework … consistently outperforms traditional methods, achieving state-of-the-art performance with an R2 of 0.8641 and an MSE of 22.41 [equating to a 14.60 per cent and 26.21 per cent increase compared to the baseline],” the authors reported in their paper.
“The success of the proposed approach suggests that it could be a valuable tool for advancing solar power generation prediction in rural areas like Alice Springs, Australia, by integrating the seasonal cycles of First Nations for improved accuracy and performance.”
Micro predictions
Predicting how much sunlight will hit a solar panel is hard: There is no universal prediction model because what reaches the Earth depends on location as much as weather and atmospheric conditions.
In 2021, South Korean researchers found a way to use machine learning to make predictions up to one day ahead, while a UNSW study in 2023 suggested that climate change is set to change the reliability of solar power in future.
But today, adding long term observational data – the kind created by Indigenous populations – can help lift the accuracy of models, the Darwin team says.
Deep learning models – those that go by the shorthand of artificial intelligence (AI) – ingest all kinds of historical data to spit out a forecast.
“In the context of Australia, various regions involve different First Nations seasonal information that reflects the diverse ecological knowledge and cultural practices of Indigenous communities throughout the country,” the paper said.
“Unlike conventional calendar systems, these seasonal insights are deeply rooted in local ecological cues, such as plant and animal behaviours, which are closely tied to changes in sunlight and weather patterns.
“By integrating this knowledge, predictions can be tailored to reflect more granular shifts in environmental conditions, leading to more precise and culturally informed forecasting for specific regions across Australia.”
Importantly, First Nation communities in northern and central Australia possess seasonal calendars that are specific to their local areas, which could be important for solar farms in remote locations.
These calendars, are the knowledge behind them could become a good framework for enhancing solar resources as well, the paper suggests.
The drawback for this particular model was that it’s tricky to build in real-time information, as the training time for the AI was “relatively long”.
