I was recently invited to be a part of the round table titled: How will new technology help to boost renewable energy integration? at the Asian Utility Week in Kuala Lumpur, Malaysia.
Before the event, I had spent some time and thought about what are the major topics that I want to share.
The round table went very well and I am really grateful to discuss these topics with such awesome people.
As I have already written down my thoughts during the preparations, I decided to share it in this blog. Below are the answers I have chosen.
What is the major issue utilities will have to overcome in order to utilize new AI technologies?
A lot of utilities are still early in the process of using new modern data-driven technologies. One of the major issues they face is data quality. Utilities majorly employ only data engineers who are responsible for the infrastructure and not for data analytics. These people usually do not analyze the data using modern data science approaches, which is the major cause why data is not preprocessed.
We have to distinguish between three data-related roles bigger internet companies usually have:
Data engineer (responsible for infrastructure, they integrate data from various resources, etc.)
Data analyst (they analyze the data using various visualizations etc.)
Data scientist (they make machine learning models etc.)
Utilities have to employ data scientists and of course, people have to start using the data in order to find all possible anomalies.
What is the Solution?
Utilities have to define their own data-driven strategy and data governance first. Last year I was at a conference about data analytics in the energy sector and the data scientist from a well known TSO told how they started their process of becoming a data-driven company. One of the first things they did is they appointed new interns to clean the data manually and throughout the time things started changing. You have to make the first step. It is always hard and time-consuming.
Another thing is to put the responsibility on the people imputing or correcting the data. If there is no responsibility people will not be careful, whereas if you put the responsibility on them they will start to care and consequently fewer mistakes will be made.
AI in the Energy Industry is Like IT 40 years ago. It started gradually, whereas nowadays there is no utility without IT department — the same will happen with AI.
What is the current “state of the art” in load forecasting?
There are a lot of journal papers published each year, everyone stating to be better than others before him. We have to keep in mind that a lot of new approachesare not validated appropriately or are validated only on a private data sets. Consequently, these “new” approaches do not generalize well to new datasets.
I think that we have to take a look at the Global Energy Forecasting Competitionswhich were organized by IEEE working group on energy forecasting in 2012, 2014 and 2017. These competitions have brought people from all around the world having a different technical background to solve load, price, solar and wind forecasting tasks. What is really interesting is that multiple linear regression based load forecasting models using appropriate feature engineering work remarkably well. Gradient boosting machines and neural network-based models are also quite popular.
Traditionally, load forecasts were required only on a transmission system level for operational purposes. In the smart grid era, a lot of technologies rely on the load forecasts on the lower levels in the network such as at a consumer level or at the transformer station level. In the smart grid era load forecasts are important from the consumer level up to a system level. The load has to be forecasted at different levels in the network, which is called hierarchical forecasting.
Another thing which is important especially at lower voltage levels in the network is providing probabilistic forecasts (PF). PFs are able to capture the uncertainty of the forecasts by providing prediction intervals. Instead of forecasting a single value (typical point forecasting), forecasts are provided in a form of quantiles. The range of prediction intervals provides uncertainty about the forecasted values. E.g. if the prediction intervals are wide (usually lower in the network) there is a bigger range of possible values and forecasts are not as certain if PI were narrower.
What is the future role of AI in the context of load forecasting?
I think that utilizing deep learning methods along with new external data sources, since deep learning is able to capture highly non-linear dependencies. Traditionally, models are based on features, which are derived from timestamps as people's behavior is highly related to the time of the day, the day of the week and the time of the year. Additionally, load highly depends on temperature which is the most important external feature and it has been widely studied in the literature.
I think that the future lies in leveraging external data sources such as traffic data for forecasting the load of the electric vehicle charging points. I think we should learn from other industries such as Uber, which are dealing with similar problems — forecasting the taxi demand.
Conclusion
Data-driven technologies are still early in the process in the energy industry. New data sources will definitely bring new challenges and I am looking forward to being a part of the change the energy industry is facing.
If you find this useful, please share this blog or join my group AI in Smart Grids on a LinkedIn and connect with others working in this field.