Using deep learning to forecast renewable energy generation
Industry
- Energy
Services
- Data Strategy
- Data Learning
Business Challenges
In a market that relies on variable energy sources such as wind it is vital to predict future energy production and how this will satisfy demand.
The National Energy Market (NEM) is a wholesale energy market for generators and retailers to trade electricity and is managed by the Australian Energy Market Operator (AEMO). The NEM includes the South Australia region, which makes heavy use of renewable energy such as wind and solar.
It is important for the NEM to be able to forecast future wind energy production in order to plan for predicted demand. The current industry standard for wind generation forecasts is not entirely transparent and relies on wind farm information that’s not publicly available. We looked at how existing models could be improved using deep learning.
Standard and consistent workflow of data science practice and the cohesion between model development and deployment were desired to accelerate the development of data science work and realise their benefits in production
Success Criteria
- Evaluate the feasibility of deep learning for predicting renewable energy output.
- Develop a real-time interactive dashboard that visualizes predicted wind energy output, actual wind energy output, and total grid demand
- Baseline deep learning solution against the Australian Wind Energy Forecasting System (AWEFS)
- A unified platform.
Solution
- Sourced historical wind generation data from AEMO/NEM, and historical weather data from the BOM.
- Established the time horizon for prediction (24 hours ahead & 1 hour increments) and establish the upper and lower bounds of prediction confidence.
- Leveraged Google Cloud’s AI Platform, and GPU accelerated compute, we iterated a range of model architectures and features using automated hyper-parameter tuning.
- Using Google Cloud Functions we download real-time data and make a new set of predictions every 5 minutes and make them available via the interactive dashboard.
Outcomes
- We evaluated performance of the model by assessing the actual wind energy output against our predicted upper and lower bounds.
- Highest accuracy achieved was 80% for predictions 1 hour out.
- The prediction tool was provided to AEMO in order to baseline performance against their existing wind energy forecasting system.