Characteristics Generation of Canadian PV Arrays by Machine Learning
Justin Du
David Thompson Secondary
Floor Location : M 094 E

Global energy generation is expected, by 2037, to be dominated by renewable energy sources, where it will comprise of more than 50% of global generation totals. Solar power in particular is an emerging technology projected to become one of the leaders of power production; predictions indicate that it could occupy nearly a third of clean energy generation in the same timeframe. Nation-wise, Canada is an outstander whose deployment of photovoltaic (PV) technologies does not match with the potential expressed by reports. Despite ranking 9th globally in PV capacity, a mere 0.5% of the total energy generation is supplied by PV technologies. Leading countries in PV potential - namely China, USA, and Japan – outshine Canada in this metric at 2.5%, 1.7%, and 6.5%.

This letdown is primarily caused by a lack of available data. Given the remoteness of a vast portion of Canada, uncertainties in long-term viability as well as lack of infrastructure restrict the speed at with PV modules can be rolled out. This is reflected once again in numbers: 98% of Canadian solar power generation is centered in Ontario, where sufficient infrastructure and data availability allow for the development of PV arrays. The long-term objective for this project aims to extend the breadth of PV modules to every region of Canada.

Machine learning (ML) may be implemented to gather the preliminary data used to determine the viability of a solar array in place of humans. This project reviews the procedures used to create two ML models that can accurately predict the first-year power production total and breakeven price of a solar array anywhere in Canada. A K-Nearest Neighbors (KNN) algorithm via scikit-learn’s KNeighborsRegressor is used in conjunction with two government published datsets to form the core model. Predictions generated by this model expressed an accuracy – measured in R-Squared (R^2) – of 76.9% and 99.8% for power production and breakeven price, respectively.

The methods explored in this project need not be limited to implementation within Canada. Unsurprisingly, the generality of these processes mean it can be applied on a global scale. “Solar-rich” locations near the equator may be targeted – especially those in which preliminary data generation is not physically possible and could instead be completed by a ML model. With large scale rollouts of PV modules, the fight against global warming can be simplified with a faster takeover of renewable energies. Machine learning, when applied in the renewable energy industry, holds the potential to simultaneously benefit both the environment and the economy. Promising accuracy scores reported in this project paint a hopeful picture for the future of this planet.