On December 20, 2018, Tokyo Electric Power Company Holdings, Inc. (TEPCO) and Hokkaido Electric Power Co., Inc. jointly launched the “PV in HOKKAIDO” contest to find outstanding new models for predicting the power output of solar power plants in Hokkaido. The models submitted by contestants were judged for their prediction accuracy.
We received around 70 entries from companies, research institutes, and students from Japan and other countries.
Thirteen contestants that submitted proposals showing high prediction accuracy, practicability, and development potential were invited to attend a final selection meeting held on June 13. The winners of the contest have now been selected.
We received proposals on a wide variety of prediction methods. One contestant, for example, achieved a high prediction accuracy using weather forecast data from the day just before the forecast date, adopting the same methodology that our company employs in actual operations. Other contestants, meanwhile, attempted to predict power output across broad areas at the same time or to predict power output without the use of weather forecast data. The contest was an advanced, sophisticated competition on the whole.
The highest-quality prediction methods proposed by the contestants will be considered for use in actual operations going forward.
We would like to thank all of the contestants for participating in this contest.
We seek to expand the PV installation and improve the stability of power supply by predicting power output with higher accuracy.
We appreciate your ongoing support and contributions.
Application Guidelines: https://cuusoo.com/projects/50369
Achieved high prediction accuracy by combining their original value prediction model and machine learning. Highly acclaimed for overall accuracy and practicality through the use of proprietary technologies to cover elements required for prediction. Their video and report were given high marks because they included a quantitative explanation of their research.
Achieved high prediction accuracy by developing a predictive method that leverages easily-obtainable prediction data from the Japan Meteorological Agency. Highly acclaimed for both the practicality and development potential of their model which can be applied to real-world tasks and easily adjusted to improve accuracy. Their video and report were also highly acclaimed for their originality and concise explanation of their research.
Achieved the highest accuracy through advanced machine learning that builds models automatically and instantly.
(See “Runner-up” for details)
Achieved high prediction accuracy by using the solar incident angle to the panel as input for machine learning.
This method is highly practical since it uses only easily-obtainable prediction data from the Japan Meteorological Agency with a combination of multiple types of machine learning.
Attempted to make a collective predition for the entire Hokkaido area. High expectations for its development potential since interesting innovations were employed, such as leveraging meteorological data for areas outside of Hokkaido.
Distinctive prediction model that is quite simple and based only on weather forecast data.
Distinctive in that the impact of fog that often occurs in the eastern part of Hokkaido is considered.
The PV power-output predictions in this contest have been performed using various methods. This contest is highly regarded for its academic value, as each of the methods applied was evaluated consistently, under the same theme. We have received a wide range of proposals for this contest, and all of them are well designed. Some contestants attempted to improve the accuracy of the prediction using conventional methods, for example, while others applied new approaches combined with machine learning. We recognize the utility of many of the proposals and also how the proposals can be further improved. We expect that the contest results will contribute to the development of better prediction methods in the future.
We expected that the predictive models for Hokkaido would include more explanatory variables than the models for other regions, as added conditions unique to the cold Hokkaido climate had to be considered. To our surprise, most of the contestants proposed methods that were not only relatively simple, but also practical in operational terms. This contest has been meaningful for the many applications it has yielded and the new methods it has helped us discover.
This contest was inspiring to me. In my opinion, the contestants who proposed conventional experience-based methods combined with new approaches ranked high. We expect to see the invention and development of prediction methods that combine conventional methods with new technologies such as AI.
New prediction methods, some of which will attain inexplicably high levels of accuracy, may emerge in step with the development of AI and other new technologies in the future. How we think of and handle these technologies will be key. The coming development should focus on two pillars: conventional methods already developed and new technologies such as AI.
This contest helped us obtain meaningful new viewpoints through the contestants’ attempts to combine various methods. While one of the proposals employs machine learning to predict power-output, the contest results show that conventional methods are still highly effective in attaining accurate predictions. We have also learned that approaches based on weather forecasts achieve more accurate predictions.
Significantly, this contest allowed us to evaluate each proposed method consistently, under the same theme. An analysis of all of the proposals submitted has provided us with interesting insights. We learned, for example, that the range of prediction accuracy can be roughly categorized by method, and that the limitations of each method can be known. We will be able to develop better methods by establishing a model integrating AI, conventional methods, and our human experience.
I have been evaluating the methods from my perspective as a worker who estimates renewable energy generation. This contest has awakened me to the different factors at work outside of my specialization. As some of proposed methods can be considered for use at a practical level, we will be seeking to make effective use of them. We believe that improved prediction accuracy will help us make more economically effective supply and demand adjustments and stimulate the wholesale electric power market overall.
The judging was held on June 13. Rankings were discussed and decided based on presentation videos and supporting references submitted by the finalists.
As the final task; top contestants that were selected based on the two submissions in the first stage were asked to turn in their presentation videos and supporting references.