This Advisian Wind and Solar Forecasting Lessons Learnt Report outlines the key learnings Advisian has learnt in the period from October 2019 – April 2020.
Advisian Wind and Solar Forecasting for the NEM Project key learnings
- Integration with Site SCADA systems requires specialist support from SCADA engineers to connect data into the Edge Gateway software. In some instances, depending on the site configurations it may require upgrade of the site hardware/software the project cost for the asset owner and the deployment lead time.
Wind Farm Forecasting Algorithm
- The power curve is non-linear with respect to wind speed so using a non-linear model is the obvious choice. However, when modelling the relationship of past/present wind power and future wind power, the choice is not so obvious and in forecasting often models that are linear in the lags and/or the past errors show good performance. We tested both linear and non linear models and concluded that a non-linear model had superior accuracy.
- During periods where the wind farm is operating under constraint, the wind power cannot be used as an input to the model. As such we have developed two models, one using Wind Power and during non-constrained periods and the other which relies on external factors such as wind speed and direction. These two models run in parallel and the algorithm switches between models depending on whether a constraint is active. This is essential for delivering an accurate forecast for when constraints are lifted.
- Extra care is needed to ensure that the algorithm normalises time zones. AEMO uses a specific time zone for forecasting which was different from the SCADA system time zone and the Edge Gateway time zones. Trending of our early models found an hour offset between the AEMO power values and SCADA wind power values which took time to debug.
- Misalignment in Units of Measure used for input variables in the model vs those received from SCADA caused erroneous forecasts which needed to be debugged. (Eg kW vs MW). These problems were solved after monitoring the input and outputs of the model over different time periods (approximately 3 hours) and comparing them with recent SCADA data.